diff --git a/egs/aishell/ASR/transducer_emformer/asr_datamodule.py b/egs/aishell/ASR/transducer_emformer/asr_datamodule.py
new file mode 120000
index 000000000..49b2ee483
--- /dev/null
+++ b/egs/aishell/ASR/transducer_emformer/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer_stateless/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_emformer/decoder.py b/egs/aishell/ASR/transducer_emformer/decoder.py
new file mode 120000
index 000000000..7f07b1a81
--- /dev/null
+++ b/egs/aishell/ASR/transducer_emformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_emformer/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_emformer/emformer.py b/egs/aishell/ASR/transducer_emformer/emformer.py
new file mode 100644
index 000000000..7e663ca32
--- /dev/null
+++ b/egs/aishell/ASR/transducer_emformer/emformer.py
@@ -0,0 +1,182 @@
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from torchaudio.models import Emformer as _Emformer
+from torchaudio.models.rnnt import _TimeReduction
+
+LOG_EPSILON = math.log(1e-10)
+
+
+class Emformer(EncoderInterface):
+    """This is just a simple wrapper around torchaudio.models.Emformer.
+    We may replace it with our own implementation some time later.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int,
+        num_encoder_layers: int,
+        segment_length: int,
+        left_context_length: int,
+        right_context_length: int,
+        max_memory_size: int = 0,
+        dropout: float = 0.1,
+        subsampling_factor: int = 4,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          segment_length:
+            Number of frames per segment.
+          left_context_length:
+            Number of frames in the left context.
+          right_context_length:
+            Number of frames in the right context.
+          max_memory_size:
+            TODO.
+          dropout:
+            Dropout in encoder.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.subsampling_factor = subsampling_factor
+
+        self.time_reduction = _TimeReduction(stride=subsampling_factor)
+        self.in_linear = nn.Linear(num_features * subsampling_factor, d_model)
+
+        self.right_context_length = right_context_length
+
+        assert right_context_length % subsampling_factor == 0
+        assert segment_length % subsampling_factor == 0
+        assert left_context_length % subsampling_factor == 0
+
+        left_context_length = left_context_length // subsampling_factor
+        right_context_length = right_context_length // subsampling_factor
+        segment_length = segment_length // subsampling_factor
+
+        self.model = _Emformer(
+            input_dim=d_model,
+            num_heads=nhead,
+            ffn_dim=dim_feedforward,
+            num_layers=num_encoder_layers,
+            segment_length=segment_length,
+            dropout=dropout,
+            activation="relu",
+            left_context_length=left_context_length,
+            right_context_length=right_context_length,
+            max_memory_size=max_memory_size,
+            weight_init_scale_strategy="depthwise",
+            tanh_on_mem=False,
+            negative_inf=-1e8,
+        )
+
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, output_dim)
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            Input features of shape (N, T, C).
+          x_lens:
+            A int32 tensor of shape (N,) containing valid frames in `x` before
+            padding. We have `x.size(1) == x_lens.max()`
+        Returns:
+          Return a tuple containing two tensors:
+
+            - encoder_out, a tensor of shape (N, T', C)
+            - encoder_out_lens, a int32 tensor of shape (N,) containing the
+              valid frames in `encoder_out` before padding
+        """
+        x = nn.functional.pad(
+            x,
+            # (left, right, top, bottom)
+            # left/right are for the channel dimension, i.e., axis 2
+            # top/bottom are for the time dimension, i.e., axis 1
+            (0, 0, 0, self.right_context_length),
+            value=LOG_EPSILON,
+        )  # (N, T, C) -> (N, T+right_context_length, C)
+
+        x, x_lens = self.time_reduction(x, x_lens)
+        x = self.in_linear(x)
+
+        emformer_out, emformer_out_lens = self.model(x, x_lens)
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens
+
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[List[List[torch.Tensor]]] = None,
+    ):
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 2-D tensor of shap containing the number of valid frames for each
+            element in `x` before padding.
+          states:
+            Internal states of the model.
+        Returns:
+          Return a tuple containing 3 tensors:
+            - encoder_out, a 3-D tensor of shape (N, T, C)
+            - encoder_out_lens: a 1-D tensor of shape (N,)
+            - next_state, internal model states for the next chunk
+        """
+        x, x_lens = self.time_reduction(x, x_lens)
+        x = self.in_linear(x)
+
+        emformer_out, emformer_out_lens, states = self.model.infer(
+            x, x_lens, states
+        )
+
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens, states
diff --git a/egs/aishell/ASR/transducer_emformer/encoder_interface.py b/egs/aishell/ASR/transducer_emformer/encoder_interface.py
new file mode 120000
index 000000000..2b42f3999
--- /dev/null
+++ b/egs/aishell/ASR/transducer_emformer/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_emformer/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_emformer/joiner.py b/egs/aishell/ASR/transducer_emformer/joiner.py
new file mode 120000
index 000000000..8cdb445b6
--- /dev/null
+++ b/egs/aishell/ASR/transducer_emformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_emformer/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_emformer/model.py b/egs/aishell/ASR/transducer_emformer/model.py
new file mode 120000
index 000000000..fbee6726a
--- /dev/null
+++ b/egs/aishell/ASR/transducer_emformer/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_emformer/model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_emformer/noam.py b/egs/aishell/ASR/transducer_emformer/noam.py
new file mode 120000
index 000000000..fd71f811e
--- /dev/null
+++ b/egs/aishell/ASR/transducer_emformer/noam.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_emformer/noam.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_emformer/test_emformer.py b/egs/aishell/ASR/transducer_emformer/test_emformer.py
new file mode 100755
index 000000000..57a1a9427
--- /dev/null
+++ b/egs/aishell/ASR/transducer_emformer/test_emformer.py
@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_emformer/test_emformer.py
+"""
+
+import warnings
+
+import torch
+from emformer import Emformer
+
+
+def test_emformer():
+    N = 3
+    T = 300
+    C = 80
+
+    subsampling_factor = 4
+
+    output_dim = 500
+
+    encoder = Emformer(
+        num_features=C,
+        output_dim=output_dim,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=20,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        subsampling_factor=4,
+    )
+
+    x = torch.rand(N, T, C)
+    x_lens = torch.randint(100, T, (N,))
+    x_lens[0] = T
+
+    y, y_lens = encoder(x, x_lens)
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        assert (y_lens == x_lens // subsampling_factor).all()
+    assert x.size(0) == x.size(0)
+    assert y.size(1) == max(y_lens)
+    assert y.size(2) == output_dim
+
+    num_param = sum([p.numel() for p in encoder.parameters()])
+    print(f"Number of encoder parameters: {num_param}")
+
+
+def main():
+    test_emformer()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220329)
+    main()
diff --git a/egs/aishell/ASR/transducer_emformer/train.py b/egs/aishell/ASR/transducer_emformer/train.py
new file mode 100755
index 000000000..8893336df
--- /dev/null
+++ b/egs/aishell/ASR/transducer_emformer/train.py
@@ -0,0 +1,851 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang
+#                                                       Mingshuang Luo)
+# Copyright    2021                           (Pingfeng Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2"
+
+./transducer_emformer/train.py \
+  --world-size 3 \
+  --num-epochs 65 \
+  --start-epoch 0 \
+  --exp-dir transducer_emformer/exp \
+  --max-duration 250 \
+  --lr-factor 2.0 \
+  --context-size 2 \
+  --modified-transducer-prob 0.25
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decoder import Decoder
+from emformer import Emformer
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from noam import Noam
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--attention-dim",
+        type=int,
+        default=512,
+        help="Attention dim for the Emformer",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads for the Emformer",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feed-forward dimension for the Emformer",
+    )
+
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of encoder layers for the Emformer",
+    )
+
+    parser.add_argument(
+        "--left-context-length",
+        type=int,
+        default=120,
+        help="Number of frames for the left context in the Emformer",
+    )
+
+    parser.add_argument(
+        "--segment-length",
+        type=int,
+        default=16,
+        help="Number of frames for each segment in the Emformer",
+    )
+
+    parser.add_argument(
+        "--right-context-length",
+        type=int,
+        default=4,
+        help="Number of frames for right context in the Emformer",
+    )
+
+    parser.add_argument(
+        "--memory-size",
+        type=int,
+        default=0,
+        help="Number of entries in the memory for the Emformer",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_emformer/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_emformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)"
+        "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 800,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for decoder
+            "embedding_dim": 512,
+            # parameters for Noam
+            "warm_step": 30000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Emformer(
+        num_features=params.feature_dim,
+        output_dim=params.vocab_size,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        left_context_length=params.left_context_length,
+        segment_length=params.segment_length,
+        right_context_length=params.right_context_length,
+        max_memory_size=params.memory_size,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.vocab_size,
+        inner_dim=params.embedding_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (
+            (feature_lens // params.subsampling_factor).sum().item()
+        )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(
+                    tb_writer, "train/tot_", params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    aishell = AishellAsrDataModule(args)
+    train_cuts = aishell.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        return 1.0 <= c.duration <= 12.0
+
+    num_in_total = len(train_cuts)
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    num_left = len(train_cuts)
+    num_removed = num_in_total - num_left
+    removed_percent = num_removed / num_in_total * 100
+
+    logging.info(f"Before removing short and long utterances: {num_in_total}")
+    logging.info(f"After removing short and long utterances: {num_left}")
+    logging.info(f"Removed {num_removed} utterances ({removed_percent:.5f}%)")
+
+    train_dl = aishell.train_dataloaders(train_cuts)
+    valid_dl = aishell.valid_dataloaders(aishell.valid_cuts())
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/learning_rate", cur_lr, params.batch_idx_train
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py
index 815e1c02a..2cb7a8cba 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py
@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import warnings
 from dataclasses import dataclass
 from typing import Dict, List, Optional
 
@@ -482,8 +483,10 @@ def modified_beam_search(
         for i in range(batch_size):
             topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
 
-            topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
-            topk_token_indexes = (topk_indexes % vocab_size).tolist()
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
 
             for k in range(len(topk_hyp_indexes)):
                 hyp_idx = topk_hyp_indexes[k]
@@ -590,8 +593,10 @@ def _deprecated_modified_beam_search(
         topk_hyp_indexes = topk_indexes // logits.size(-1)
         topk_token_indexes = topk_indexes % logits.size(-1)
 
-        topk_hyp_indexes = topk_hyp_indexes.tolist()
-        topk_token_indexes = topk_token_indexes.tolist()
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            topk_hyp_indexes = topk_hyp_indexes.tolist()
+            topk_token_indexes = topk_token_indexes.tolist()
 
         for i in range(len(topk_hyp_indexes)):
             hyp = A[topk_hyp_indexes[i]]
diff --git a/egs/librispeech/ASR/transducer_emformer/__init__.py b/egs/librispeech/ASR/transducer_emformer/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/transducer_emformer/asr_datamodule.py b/egs/librispeech/ASR/transducer_emformer/asr_datamodule.py
new file mode 120000
index 000000000..b4e5427e0
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_emformer/beam_search.py b/egs/librispeech/ASR/transducer_emformer/beam_search.py
new file mode 120000
index 000000000..227d2247c
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_emformer/decode.py b/egs/librispeech/ASR/transducer_emformer/decode.py
new file mode 100755
index 000000000..c40b01dfa
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/decode.py
@@ -0,0 +1,549 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./transducer_emformer/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./transducer_emformer/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./transducer_emformer/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./transducer_emformer/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./transducer_emformer/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./transducer_emformer/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./transducer_emformer/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./transducer_emformer/exp \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--avg-last-n",
+        type=int,
+        default=0,
+        help="""If positive, --epoch and --avg are ignored and it
+        will use the last n checkpoints exp_dir/checkpoint-xxx.pt
+        where xxx is the number of processed batches while
+        saving that checkpoint.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_emformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=feature, x_lens=feature_lens
+    )
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif (
+        params.decoding_method == "greedy_search"
+        and params.max_sym_per_frame == 1
+    ):
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for hyp_words, ref_text in zip(hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(
+                f"batch {batch_str}, cuts processed until now is {num_cuts}"
+            )
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir
+        / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg_last_n > 0:
+        filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_emformer/decoder.py b/egs/librispeech/ASR/transducer_emformer/decoder.py
new file mode 120000
index 000000000..0d5f10dc0
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_emformer/emformer.py b/egs/librispeech/ASR/transducer_emformer/emformer.py
new file mode 100644
index 000000000..b3693d660
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/emformer.py
@@ -0,0 +1,200 @@
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import warnings
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torchaudio.models import Emformer as _Emformer
+
+LOG_EPSILON = math.log(1e-10)
+
+
+class Emformer(EncoderInterface):
+    """This is just a simple wrapper around torchaudio.models.Emformer.
+    We may replace it with our own implementation some time later.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int,
+        num_encoder_layers: int,
+        segment_length: int,
+        left_context_length: int,
+        right_context_length: int,
+        max_memory_size: int = 0,
+        dropout: float = 0.1,
+        subsampling_factor: int = 4,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          segment_length:
+            Number of frames per segment.
+          left_context_length:
+            Number of frames in the left context.
+          right_context_length:
+            Number of frames in the right context.
+          max_memory_size:
+            TODO.
+          dropout:
+            Dropout in encoder.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.right_context_length = right_context_length
+
+        assert right_context_length % subsampling_factor == 0
+        assert segment_length % subsampling_factor == 0
+        assert left_context_length % subsampling_factor == 0
+
+        left_context_length = left_context_length // subsampling_factor
+        right_context_length = right_context_length // subsampling_factor
+        segment_length = segment_length // subsampling_factor
+
+        self.model = _Emformer(
+            input_dim=d_model,
+            num_heads=nhead,
+            ffn_dim=dim_feedforward,
+            num_layers=num_encoder_layers,
+            segment_length=segment_length,
+            dropout=dropout,
+            activation="relu",
+            left_context_length=left_context_length,
+            right_context_length=right_context_length,
+            max_memory_size=max_memory_size,
+            weight_init_scale_strategy="depthwise",
+            tanh_on_mem=False,
+            negative_inf=-1e8,
+        )
+
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, output_dim)
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            Input features of shape (N, T, C).
+          x_lens:
+            A int32 tensor of shape (N,) containing valid frames in `x` before
+            padding. We have `x.size(1) == x_lens.max()`
+        Returns:
+          Return a tuple containing two tensors:
+
+            - encoder_out, a tensor of shape (N, T', C)
+            - encoder_out_lens, a int32 tensor of shape (N,) containing the
+              valid frames in `encoder_out` before padding
+        """
+        x = nn.functional.pad(
+            x,
+            # (left, right, top, bottom)
+            # left/right are for the channel dimension, i.e., axis 2
+            # top/bottom are for the time dimension, i.e., axis 1
+            (0, 0, 0, self.right_context_length),
+            value=LOG_EPSILON,
+        )  # (N, T, C) -> (N, T+right_context_length, C)
+
+        x = self.encoder_embed(x)
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            # Caution: We assume the subsampling factor is 4!
+            x_lens = ((x_lens - 1) // 2 - 1) // 2
+
+        emformer_out, emformer_out_lens = self.model(x, x_lens)
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens
+
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[List[List[torch.Tensor]]] = None,
+    ):
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 2-D tensor of shap containing the number of valid frames for each
+            element in `x` before padding.
+          states:
+            Internal states of the model.
+        Returns:
+          Return a tuple containing 3 tensors:
+            - encoder_out, a 3-D tensor of shape (N, T, C)
+            - encoder_out_lens: a 1-D tensor of shape (N,)
+            - next_state, internal model states for the next chunk
+        """
+        x = self.encoder_embed(x)
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            # Caution: We assume the subsampling factor is 4!
+            x_lens = ((x_lens - 1) // 2 - 1) // 2
+        emformer_out, emformer_out_lens, states = self.model.infer(
+            x, x_lens, states
+        )
+
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens, states
diff --git a/egs/librispeech/ASR/transducer_emformer/encoder_interface.py b/egs/librispeech/ASR/transducer_emformer/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_emformer/joiner.py b/egs/librispeech/ASR/transducer_emformer/joiner.py
new file mode 120000
index 000000000..81ad47c55
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_emformer/model.py b/egs/librispeech/ASR/transducer_emformer/model.py
new file mode 120000
index 000000000..a61a0a23f
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_emformer/noam.py b/egs/librispeech/ASR/transducer_emformer/noam.py
new file mode 100644
index 000000000..e46bf35fb
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/noam.py
@@ -0,0 +1,104 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
diff --git a/egs/librispeech/ASR/transducer_emformer/streaming_decode.py b/egs/librispeech/ASR/transducer_emformer/streaming_decode.py
new file mode 100755
index 000000000..0b21b3f59
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/streaming_decode.py
@@ -0,0 +1,362 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import time
+from pathlib import Path
+from typing import List, Optional
+
+import kaldifeat
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import AttributeDict, setup_logger
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--avg-last-n",
+        type=int,
+        default=0,
+        help="""If positive, --epoch and --avg are ignored and it
+        will use the last n checkpoints exp_dir/checkpoint-xxx.pt
+        where xxx is the number of processed batches while
+        saving that checkpoint.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_emformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_feature_extractor(
+    params: AttributeDict,
+) -> kaldifeat.Fbank:
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = params.device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = True
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    return kaldifeat.Fbank(opts)
+
+
+def decode_one_utterance(
+    audio_samples: torch.Tensor,
+    model: nn.Module,
+    fbank: kaldifeat.Fbank,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+):
+    """Decode one utterance.
+    Args:
+      audio_samples:
+        A 1-D float32 tensor of shape (num_samples,) containing the normalized
+        audio samples. Normalized means the samples is in the range [-1, 1].
+      model:
+        The RNN-T model.
+      feature_extractor:
+        The feature extractor.
+      params:
+        It is the return value of :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    sample_rate = params.sample_rate
+    frame_shift = sample_rate * fbank.opts.frame_opts.frame_shift_ms / 1000
+
+    frame_shift = int(frame_shift)  # number of samples
+
+    # Note: We add 3 here because the subsampling method ((n-1)//2-1))//2
+    # is not equal to n//4. We will switch to a subsampling method that
+    # satisfies n//4, where n is the number of input frames.
+    segment_length = (params.segment_length + 3) * frame_shift
+
+    right_context_length = params.right_context_length * frame_shift
+    chunk_size = segment_length + right_context_length
+
+    opts = fbank.opts.frame_opts
+    chunk_size += (
+        (opts.frame_length_ms - opts.frame_shift_ms) / 1000 * sample_rate
+    )
+
+    chunk_size = int(chunk_size)
+
+    states: Optional[List[List[torch.Tensor]]] = None
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    hyp = [blank_id] * context_size
+
+    decoder_input = torch.tensor(hyp, device=device, dtype=torch.int64).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    i = 0
+    num_samples = audio_samples.size(0)
+    while i < num_samples:
+        # Note: The current approach of computing the features is not ideal
+        # since it re-computes the features for the right context.
+        chunk = audio_samples[i : i + chunk_size]  # noqa
+        i += segment_length
+        if chunk.size(0) < chunk_size:
+            chunk = torch.nn.functional.pad(
+                chunk, pad=(0, chunk_size - chunk.size(0))
+            )
+        features = fbank(chunk)
+        feature_lens = torch.tensor([features.size(0)], device=params.device)
+
+        features = features.unsqueeze(0)  # (1, T, C)
+
+        encoder_out, encoder_out_lens, states = model.encoder.streaming_forward(
+            features,
+            feature_lens,
+            states,
+        )
+        for t in range(encoder_out_lens.item()):
+            # fmt: off
+            current_encoder_out = encoder_out[0:1, t:t+1, :].unsqueeze(2)
+            # fmt: on
+            logits = model.joiner(current_encoder_out, decoder_out.unsqueeze(1))
+            # logits is (1, 1, 1, vocab_size)
+            y = logits.argmax().item()
+            if y == blank_id:
+                continue
+
+            hyp.append(y)
+
+            decoder_input = torch.tensor(
+                [hyp[-context_size:]], device=device, dtype=torch.int64
+            ).reshape(1, context_size)
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+        logging.info(f"Partial result:\n{sp.decode(hyp[context_size:])}")
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    # Note: params.decoding_method is currently not used.
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    setup_logger(f"{params.res_dir}/log-streaming-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    params.device = device
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg_last_n > 0:
+        filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+
+    fbank = get_feature_extractor(params)
+
+    for num, cut in enumerate(test_clean_cuts):
+        logging.info("Processing {num}")
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        decode_one_utterance(
+            audio_samples=torch.from_numpy(audio).squeeze(0).to(device),
+            model=model,
+            fbank=fbank,
+            params=params,
+            sp=sp,
+        )
+
+        logging.info(f"The ground truth is:\n{cut.supervisions[0].text}")
+        if num >= 0:
+            break
+        time.sleep(2)  # So that you can see the decoded results
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_emformer/subsampling.py b/egs/librispeech/ASR/transducer_emformer/subsampling.py
new file mode 120000
index 000000000..6fee09e58
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/subsampling.py
@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_emformer/test_emformer.py b/egs/librispeech/ASR/transducer_emformer/test_emformer.py
new file mode 100755
index 000000000..d8c7b37e2
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/test_emformer.py
@@ -0,0 +1,74 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_emformer/test_emformer.py
+"""
+
+import warnings
+
+import torch
+from emformer import Emformer
+
+
+def test_emformer():
+    N = 3
+    T = 300
+    C = 80
+
+    output_dim = 500
+
+    encoder = Emformer(
+        num_features=C,
+        output_dim=output_dim,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=20,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+
+    x = torch.rand(N, T, C)
+    x_lens = torch.randint(100, T, (N,))
+    x_lens[0] = T
+
+    y, y_lens = encoder(x, x_lens)
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        assert (y_lens == ((x_lens - 1) // 2 - 1) // 2).all()
+    assert x.size(0) == x.size(0)
+    assert y.size(1) == max(y_lens)
+    assert y.size(2) == output_dim
+
+    num_param = sum([p.numel() for p in encoder.parameters()])
+    print(f"Number of encoder parameters: {num_param}")
+
+
+def main():
+    test_emformer()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220329)
+    main()
diff --git a/egs/librispeech/ASR/transducer_emformer/train.py b/egs/librispeech/ASR/transducer_emformer/train.py
new file mode 100755
index 000000000..9798fe5e6
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_emformer/train.py
@@ -0,0 +1,998 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_emformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_emformer/exp \
+  --full-libri 1 \
+  --max-duration 300
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from emformer import Emformer
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from noam import Noam
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    measure_gradient_norms,
+    measure_weight_norms,
+    optim_step_and_measure_param_change,
+    setup_logger,
+    str2bool,
+)
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--attention-dim",
+        type=int,
+        default=512,
+        help="Attention dim for the Emformer",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads for the Emformer",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feed-forward dimension for the Emformer",
+    )
+
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of encoder layers for the Emformer",
+    )
+
+    parser.add_argument(
+        "--left-context-length",
+        type=int,
+        default=120,
+        help="Number of frames for the left context in the Emformer",
+    )
+
+    parser.add_argument(
+        "--segment-length",
+        type=int,
+        default=16,
+        help="Number of frames for each segment in the Emformer",
+    )
+
+    parser.add_argument(
+        "--right-context-length",
+        type=int,
+        default=4,
+        help="Number of frames for right context in the Emformer",
+    )
+
+    parser.add_argument(
+        "--memory-size",
+        type=int,
+        default=0,
+        help="Number of entries in the memory for the Emformer",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_emformer/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_emformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)"
+        "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            "log_diagnostics": False,
+            # parameters for Emformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            # parameters for decoder
+            "embedding_dim": 512,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 20000
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Emformer(
+        num_features=params.feature_dim,
+        output_dim=params.vocab_size,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        left_context_length=params.left_context_length,
+        segment_length=params.segment_length,
+        right_context_length=params.right_context_length,
+        max_memory_size=params.memory_size,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.vocab_size,
+        inner_dim=params.embedding_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    sampler: Optional[CutSampler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        sampler=sampler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Emformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (
+            (feature_lens // params.subsampling_factor).sum().item()
+        )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    def maybe_log_gradients(tag: str):
+        if (
+            params.log_diagnostics
+            and tb_writer is not None
+            and params.batch_idx_train % (params.log_interval * 5) == 0
+        ):
+            tb_writer.add_scalars(
+                tag,
+                measure_gradient_norms(model, norm="l2"),
+                global_step=params.batch_idx_train,
+            )
+
+    def maybe_log_weights(tag: str):
+        if (
+            params.log_diagnostics
+            and tb_writer is not None
+            and params.batch_idx_train % (params.log_interval * 5) == 0
+        ):
+            tb_writer.add_scalars(
+                tag,
+                measure_weight_norms(model, norm="l2"),
+                global_step=params.batch_idx_train,
+            )
+
+    def maybe_log_param_relative_changes():
+        if (
+            params.log_diagnostics
+            and tb_writer is not None
+            and params.batch_idx_train % (params.log_interval * 5) == 0
+        ):
+            deltas = optim_step_and_measure_param_change(model, optimizer)
+            tb_writer.add_scalars(
+                "train/relative_param_change_per_minibatch",
+                deltas,
+                global_step=params.batch_idx_train,
+            )
+        else:
+            optimizer.step()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        loss.backward()
+
+        maybe_log_weights("train/param_norms")
+        maybe_log_gradients("train/grad_norms")
+        maybe_log_param_relative_changes()
+
+        optimizer.zero_grad()
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                sampler=train_dl.sampler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(
+                    tb_writer, "train/tot_", params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+        params.warm_step = 20000
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        train_cuts += librispeech.train_clean_360_cuts()
+        train_cuts += librispeech.train_other_500_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    num_in_total = len(train_cuts)
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    num_left = len(train_cuts)
+    num_removed = num_in_total - num_left
+    removed_percent = num_removed / num_in_total * 100
+
+    logging.info(f"Before removing short and long utterances: {num_in_total}")
+    logging.info(f"After removing short and long utterances: {num_left}")
+    logging.info(f"Removed {num_removed} utterances ({removed_percent:.5f}%)")
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        sp=sp,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/learning_rate", cur_lr, params.batch_idx_train
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sampler=train_dl.sampler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()