update documentation for PromptASR

egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py

@@ -211,7 +211,7 @@ class Aidatatang_200zhAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/aishell/ASR/RESULTS.md

@@ -2,6 +2,56 @@
 
 ### Aishell training result(Stateless Transducer)
 
+#### Pruned transducer stateless 7 streaming
+[./pruned_transducer_stateless7_streaming](./pruned_transducer_stateless7_streaming)
+
+It's Streaming version of Zipformer1 with Pruned RNNT loss.
+
+|                        | test | dev  | comment                               |
+|------------------------|------|------|---------------------------------------|
+| greedy search          | 6.95 | 6.29 | --epoch 44 --avg 15 --max-duration 600 |
+| modified beam search   | 6.51 | 5.90 | --epoch 44 --avg 15 --max-duration 600 |
+| fast beam search       | 6.73 | 6.09 | --epoch 44 --avg 15 --max-duration 600 |
+
+Training command is:
+
+```bash
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./pruned_transducer_stateless7_streaming/train.py \
+    --world-size 2 \
+    --num-epochs 50 \
+    --use-fp16 1 \
+    --context-size 1 \
+    --max-duration 800 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --enable-musan 0 \
+    --spec-aug-time-warp-factor 20
+```
+
+**Caution**: It uses `--context-size=1`.
+
+The decoding command is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search ; do
+  ./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 44 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --lang-dir data/lang_char \
+    --context-size 1 \
+    --decoding-method $m
+done
+```
+
+Pretrained models, training logs, decoding logs, tensorboard and decoding results
+are available at
+<https://huggingface.co/zrjin/icefall-asr-aishell-zipformer-pruned-transducer-stateless7-streaming-2023-10-16/>
+
+
+
 #### Pruned transducer stateless 7
 
 [./pruned_transducer_stateless7](./pruned_transducer_stateless7)

egs/aishell/ASR/prepare.sh

@@ -204,10 +204,6 @@ if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
   if [ ! -f $lang_char_dir/L_disambig.pt ]; then
     ./local/prepare_char.py --lang-dir $lang_char_dir
   fi
-
-  if [ ! -f $lang_char_dir/HLG.fst ]; then
-    ./local/prepare_lang_fst.py  --lang-dir $lang_phone_dir --ngram-G ./data/lm/G_3_gram.fst.txt
-  fi
 fi
 
 if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
@@ -262,6 +258,13 @@ if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
       --max-order=3 \
       data/lm/3-gram.unpruned.arpa > data/lm/G_3_gram_char.fst.txt
   fi
+
+  if [ ! -f $lang_char_dir/HLG.fst ]; then
+    lang_phone_dir=data/lang_phone
+    ./local/prepare_lang_fst.py  \
+      --lang-dir $lang_phone_dir \
+      --ngram-G ./data/lm/G_3_gram.fst.txt
+  fi
 fi
 
 if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then

egs/aishell/ASR/pruned_transducer_stateless7/train.py

@@ -703,7 +703,7 @@ def compute_loss(
             if batch_idx_train >= warm_step
             else 0.1 + 0.9 * (batch_idx_train / warm_step)
         )
-        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
 
     assert loss.requires_grad == is_training
 

egs/aishell/ASR/pruned_transducer_stateless7_streaming/README.md

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/README.md

egs/aishell/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/beam_search.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py

@@ -0,0 +1,735 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# 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
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import ContextGraph
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    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 integer 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=1,
+        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(
+        "--context-score",
+        type=float,
+        default=2,
+        help="""
+        The bonus score of each token for the context biasing words/phrases.
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-file",
+        type=str,
+        default="",
+        help="""
+        The path of the context biasing lists, one word/phrase each line
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = 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.
+      token_table:
+        It maps token ID to a string.
+      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 = next(model.parameters()).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)
+
+    feature_lens += 30
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, 30),
+        value=LOG_EPS,
+    )
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            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,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            context_graph=context_graph,
+        )
+    else:
+        hyp_tokens = []
+        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}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    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:
+        key = f"beam_size_{params.beam_size}"
+        if params.has_contexts:
+            key += f"-context-score-{params.context_score}"
+        else:
+            key += "-no-context-words"
+        return {key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[Tuple[str, 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.
+      token_table:
+        It maps a token ID to a string.
+      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 = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, 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[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        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}-{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}-{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()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+
+    if os.path.exists(params.context_file):
+        params.has_contexts = True
+    else:
+        params.has_contexts = False
+
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        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"-{params.decoding_method}-beam-size-{params.beam_size}"
+        if params.has_contexts:
+            params.suffix += f"-context-score-{params.context_score}"
+        else:
+            params.suffix += "-no-contexts-words"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    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}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            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 i >= 1:
+                    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))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    if params.decoding_method == "modified_beam_search":
+        if os.path.exists(params.context_file):
+            contexts_text = []
+            for line in open(params.context_file).readlines():
+                contexts_text.append(line.strip())
+            contexts = graph_compiler.texts_to_ids(contexts_text)
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build(contexts)
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+
+    test_cuts = aishell.test_cuts()
+    dev_cuts = aishell.valid_cuts()
+
+    test_dl = aishell.test_dataloaders(test_cuts)
+    dev_dl = aishell.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    import time
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        start = time.time()
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+        )
+        logging.info(f"Elasped time for {test_set}: {time.time() - start}")
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode_stream.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/decoder.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/export.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/joiner.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/model.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/model.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_check.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_check.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/optim.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/pretrained.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/streaming_beam_search.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py

@@ -0,0 +1,627 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, 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:
+./pruned_transducer_stateless7_streaming/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --decode-chunk-len 32 \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+import os
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import stack_states, unstack_states
+
+from icefall import ContextGraph
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+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 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is 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=32,
+        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(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
+    # factor in encoders is 8.
+    # After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
+    tail_length = 23
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset 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.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = model.encoder.get_init_state(device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        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
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        [
+                            token_table[result]
+                            for result in decode_streams[i].decoding_result()
+                        ],
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    [
+                        token_table[result]
+                        for result in decode_streams[i].decoding_result()
+                    ],
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        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}-{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}-{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()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    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}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            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))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    if params.decoding_method == "modified_beam_search":
+        if os.path.exists(params.context_file):
+            contexts_text = []
+            for line in open(params.context_file).readlines():
+                contexts_text.append(line.strip())
+            contexts = graph_compiler.texts_to_ids(contexts_text)
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build(contexts)
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    aishell = AishellAsrDataModule(args)
+
+    test_cuts = aishell.test_cuts()
+    valid_cuts = aishell.valid_cuts()
+
+    test_sets = ["test", "valid"]
+    cuts = [test_cuts, valid_cuts]
+
+    for test_set, test_cut in zip(test_sets, cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/aishell/ASR/pruned_transducer_stateless7_streaming/test_model.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/test_model.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py

egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer2.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer2.py

egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py

@@ -160,7 +160,7 @@ class AsrDataModule:
         if cuts_musan is not None:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py

@@ -218,7 +218,7 @@ class AiShell2AsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py

@@ -228,7 +228,7 @@ class Aishell4AsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py

@@ -211,7 +211,7 @@ class AlimeetingAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/alimeeting/ASR_v2/pruned_transducer_stateless7/asr_datamodule.py

@@ -208,7 +208,7 @@ class AlimeetingAsrDataModule:
             logging.info("Enable MUSAN")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
@@ -288,7 +288,6 @@ class AlimeetingAsrDataModule:
         return train_dl
 
     def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
-
         transforms = []
         if self.args.concatenate_cuts:
             transforms = [

egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py

@@ -214,7 +214,7 @@ class AmiAsrDataModule:
             logging.info("Enable MUSAN")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/ami/SURT/dprnn_zipformer/asr_datamodule.py

@@ -202,7 +202,7 @@ class AmiAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py

@@ -230,7 +230,7 @@ class CommonVoiceAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/csj/ASR/local/utils/asr_datamodule.py

@@ -256,7 +256,7 @@ class CSJAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.musan_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py

@@ -194,7 +194,7 @@ class GigaSpeechAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py

@@ -217,7 +217,7 @@ class GigaSpeechAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py

@@ -204,7 +204,7 @@ class LibriCssAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/libricss/SURT/prepare.sh

@@ -90,6 +90,14 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
   # NOTE: Alignments are required for this recipe.
   mkdir -p data/manifests
 
+  log "This recipe uses mfa alignment for trimming"
+  if [ ! -d $dl_dir/libri_alignments/LibriSpeech ]; then
+    log "No alignment provided. please refer to ../../librispeech/ASR/add_alignments.sh \n \
+        for mfa alignments. Once you have downloaded and unzipped the .zip file containing \n \
+        all alignments, the folder should be renamed to libri_alignments and moved to your $dl_dir ."
+    exit 0
+  fi
+
   lhotse prepare librispeech -p train-clean-100 -p train-clean-360 -p train-other-500 -p dev-clean \
     -j 4 --alignments-dir $dl_dir/libri_alignments/LibriSpeech $dl_dir/librispeech data/manifests/
 fi
@@ -118,9 +126,12 @@ fi
 
 if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
   log "Stage 4: Extract features for LibriSpeech, trim to alignments, and shuffle the cuts"
-  python local/compute_fbank_librispeech.py
-  lhotse combine data/manifests/librispeech_cuts_train* - |\
-    lhotse cut trim-to-alignments --type word --max-pause 0.2 - - |\
+  # python local/compute_fbank_librispeech.py
+  lhotse combine data/manifests/librispeech_cuts_train* data/manifests/librispeech_cuts_train_all.jsonl.gz
+  lhotse cut trim-to-alignments --type word --max-pause 0.2 \
+    data/manifests/librispeech_cuts_train_all.jsonl.gz \
+    data/manifests/librispeech_cuts_train_all_trimmed.jsonl.gz
+  cat <(gunzip -c data/manifests/librispeech_cuts_train_all_trimmed.jsonl.gz) | \
     shuf | gzip -c > data/manifests/librispeech_cuts_train_trimmed.jsonl.gz
 fi
 
@@ -152,7 +163,7 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
     data/manifests/lsmix_cuts_train_clean_ov40.jsonl.gz
 
   # Full training set (2,3 speakers) anechoic
-  log "Generating anechoic ${part} set (full)"
+  log "Generating anechoic set (full)"
   lhotse workflows simulate-meetings \
     --method conversational \
     --fit-to-supervisions data/manifests/libricss-sdm_supervisions_all_v1.jsonl.gz \

egs/librispeech/ASR/prepare.sh

@@ -242,10 +242,6 @@ if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
         $lang_dir/L_disambig.pt \
         $lang_dir/L_disambig.fst
     fi
-
-    if [ ! -f $lang_dir/HL.fst ]; then
-      ./local/prepare_lang_fst.py  --lang-dir $lang_dir --ngram-G ./data/lm/G_3_gram.fst.txt
-    fi
   done
 fi
 
@@ -303,6 +299,16 @@ if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
       --max-order=4 \
       $dl_dir/lm/4-gram.arpa > data/lm/G_4_gram.fst.txt
   fi
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/HL.fst ]; then
+      ./local/prepare_lang_fst.py  \
+        --lang-dir $lang_dir \
+        --ngram-G ./data/lm/G_3_gram.fst.txt
+    fi
+  done
 fi
 
 if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then

egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py

@@ -209,7 +209,7 @@ class LibriSpeechAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "cuts_musan.json.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py

@@ -158,7 +158,7 @@ class Conformer(EncoderInterface):
         if not is_jit_tracing():
             assert x.size(0) == lengths.max().item()
 
-        src_key_padding_mask = make_pad_mask(lengths)
+        src_key_padding_mask = make_pad_mask(lengths, x.size(0))
 
         if self.dynamic_chunk_training:
             assert (

egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py

@@ -164,7 +164,7 @@ class AsrDataModule:
         if cuts_musan is not None:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py

@@ -217,7 +217,7 @@ class GigaSpeechAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py

@@ -281,7 +281,7 @@ class Zipformer(EncoderInterface):
 
         lengths = (x_lens - 7) >> 1
         assert x.size(0) == lengths.max().item(), (x.shape, lengths, lengths.max())
-        mask = make_pad_mask(lengths)
+        mask = make_pad_mask(lengths, x.size(0))
 
         outputs = []
         feature_masks = self.get_feature_masks(x)

egs/librispeech/ASR/pruned_transducer_stateless8/export-onnx.py

@@ -0,0 +1,604 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/WeijiZhuang/icefall-asr-librispeech-pruned-transducer-stateless8-2022-12-02
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/WeijiZhuang/icefall-asr-librispeech-pruned-transducer-stateless8-2022-12-02
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless8/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --feedforward-dims "1024,1024,2048,2048,1024"
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import Zipformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+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 averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Zipformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless7",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # <blk> is defined in local/train_bpe_model.py
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        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 i >= 1:
+                    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), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()

egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py

@@ -233,7 +233,7 @@ class LibriSpeechAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/librispeech/ASR/zipformer/zipformer.py

@@ -17,28 +17,33 @@
 # limitations under the License.
 
 import copy
+import logging
 import math
+import random
 import warnings
 from typing import List, Optional, Tuple, Union
-import logging
+
 import torch
-import random
 from encoder_interface import EncoderInterface
 from scaling import (
+    Identity,  # more friendly to backward hooks than nn.Identity(), for diagnostic reasons.
+)
+from scaling import (
+    ScaledLinear,  # not as in other dirs.. just scales down initial parameter values.
+)
+from scaling import (
+    ActivationDropoutAndLinear,
     Balancer,
     BiasNorm,
-    Dropout2,
     ChunkCausalDepthwiseConv1d,
-    ActivationDropoutAndLinear,
-    ScaledLinear,  # not as in other dirs.. just scales down initial parameter values.
+    Dropout2,
+    FloatLike,
+    ScheduledFloat,
     Whiten,
-    Identity,  # more friendly to backward hooks than nn.Identity(), for diagnostic reasons.
+    convert_num_channels,
+    limit_param_value,
     penalize_abs_values_gt,
     softmax,
-    ScheduledFloat,
-    FloatLike,
-    limit_param_value,
-    convert_num_channels,
 )
 from torch import Tensor, nn
 
@@ -2098,7 +2103,7 @@ class NonlinAttention(nn.Module):
         (seq_len, batch_size, _) = x.shape
         hidden_channels = self.hidden_channels
 
-        s, x, y = x.chunk(3, dim=-1)
+        s, x, y = x.chunk(3, dim=2)
 
         # s will go through tanh.
 
@@ -2151,7 +2156,7 @@ class NonlinAttention(nn.Module):
         (seq_len, batch_size, _) = x.shape
         hidden_channels = self.hidden_channels
 
-        s, x, y = x.chunk(3, dim=-1)
+        s, x, y = x.chunk(3, dim=2)
 
         # s will go through tanh.
         s = self.tanh(s)
@@ -2308,7 +2313,7 @@ class ConvolutionModule(nn.Module):
 
         x = self.in_proj(x)  # (time, batch, 2*channels)
 
-        x, s = x.chunk(2, dim=-1)
+        x, s = x.chunk(2, dim=2)
         s = self.balancer1(s)
         s = self.sigmoid(s)
         x = self.activation1(x)  # identity.

egs/librispeech/WSASR/conformer_ctc2/asr_datamodule.py

@@ -32,7 +32,7 @@ from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
     DynamicBucketingSampler,
     K2SpeechRecognitionDataset,
     PrecomputedFeatures,
-    SingleCutSampler,
+    SimpleCutSampler,
     SpecAugment,
 )
 from lhotse.dataset.input_strategies import AudioSamples  # noqa F401 For AudioSamples
@@ -230,8 +230,8 @@ class LibriSpeechAsrDataModule:
                 drop_last=self.args.drop_last,
             )
         else:
-            logging.info("Using SingleCutSampler.")
-            train_sampler = SingleCutSampler(
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
                 cuts_train,
                 max_duration=self.args.max_duration,
                 shuffle=self.args.shuffle,

egs/mgb2/ASR/conformer_ctc/asr_datamodule.py

@@ -182,7 +182,6 @@ class MGB2AsrDataModule:
         cuts_train: CutSet,
         sampler_state_dict: Optional[Dict[str, Any]] = None,
     ) -> DataLoader:
-
         transforms = []
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
@@ -190,7 +189,7 @@ class MGB2AsrDataModule:
             cuts_musan = load_manifest(self.args.manifest_dir / "cuts_musan.jsonl.gz")
 
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py

@@ -219,7 +219,7 @@ class AsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/spgispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py

@@ -182,7 +182,7 @@ class SPGISpeechAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
@@ -261,7 +261,6 @@ class SPGISpeechAsrDataModule:
         return train_dl
 
     def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
-
         transforms = []
         if self.args.concatenate_cuts:
             transforms = [

egs/swbd/ASR/conformer_ctc/asr_datamodule.py

@@ -32,7 +32,7 @@ from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
     DynamicBucketingSampler,
     K2SpeechRecognitionDataset,
     PrecomputedFeatures,
-    SingleCutSampler,
+    SimpleCutSampler,
     SpecAugment,
 )
 from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
@@ -220,7 +220,7 @@ class SwitchBoardAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
@@ -302,8 +302,8 @@ class SwitchBoardAsrDataModule:
                 buffer_size=50000,
             )
         else:
-            logging.info("Using SingleCutSampler.")
-            train_sampler = SingleCutSampler(
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
                 cuts_train,
                 max_duration=self.args.max_duration,
                 shuffle=self.args.shuffle,

egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py

@@ -228,7 +228,7 @@ class TAL_CSASRAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py

@@ -205,7 +205,7 @@ class TedLiumAsrDataModule:
             logging.info("Enable MUSAN")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py

@@ -157,7 +157,7 @@ class TimitAsrDataModule(DataModule):
         cuts_musan = load_manifest(self.args.feature_dir / "musan_cuts.jsonl.gz")
 
         logging.info("About to create train dataset")
-        transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))]
+        transforms = [CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20))]
         if self.args.concatenate_cuts:
             logging.info(
                 f"Using cut concatenation with duration factor "

egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py

@@ -215,7 +215,7 @@ class WenetSpeechAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")

egs/wenetspeech/ASR/pruned_transducer_stateless2/finetune.py

@@ -734,7 +734,7 @@ def train_one_epoch(
             scaler.update()
             optimizer.zero_grad()
         except:  # noqa
-            display_and_save_batch(batch, params=params)
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
             raise
 
         if params.print_diagnostics and batch_idx == 5:

egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py

@@ -218,7 +218,7 @@ class Xbmu_AmdoAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")