Add non-streaming Zipformer recipe for KsponSpeech (#1664)

egs/ksponspeech/ASR/README.md

@@ -28,5 +28,6 @@ There are various folders containing the name `transducer` in this folder. The f
 |                                          | Encoder              | Decoder            | Comment                                           |
 | ---------------------------------------- | -------------------- | ------------------ | ------------------------------------------------- |
 | `pruned_transducer_stateless7_streaming` | Streaming Zipformer  | Embedding + Conv1d | streaming version of pruned_transducer_stateless7 |
+| `zipformer`                              | Upgraded Zipformer   | Embedding + Conv1d | The latest recipe                                 |
 
 The decoder in `transducer_stateless` is modified from the paper [Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/). We place an additional Conv1d layer right after the input embedding layer.

egs/ksponspeech/ASR/RESULTS.md

@@ -19,13 +19,13 @@ The CERs are:
 | fast beam search     | 320ms      | 10.21      | 11.04      | --epoch 30 --avg 9  | simulated streaming  |
 | fast beam search     | 320ms      | 10.25      | 11.08      | --epoch 30 --avg 9  | chunk-wise           |
 | modified beam search | 320ms      | 10.13      | 10.88      | --epoch 30 --avg 9  | simulated streaming  |
-| modified beam search | 320ms      | 10.1       | 10.93      | --epoch 30 --avg 9  | chunk-size           |
+| modified beam search | 320ms      | 10.1       | 10.93      | --epoch 30 --avg 9  | chunk-wize           |
 | greedy search        | 640ms      | 9.94       | 10.82      | --epoch 30 --avg 9  | simulated streaming  |
 | greedy search        | 640ms      | 10.04      | 10.85      | --epoch 30 --avg 9  | chunk-wise           |
 | fast beam search     | 640ms      | 10.01      | 10.81      | --epoch 30 --avg 9  | simulated streaming  |
 | fast beam search     | 640ms      | 10.04      | 10.7       | --epoch 30 --avg 9  | chunk-wise           |
 | modified beam search | 640ms      | 9.91       | 10.72      | --epoch 30 --avg 9  | simulated streaming  |
-| modified beam search | 640ms      | 9.92       | 10.72      | --epoch 30 --avg 9  | chunk-size           |
+| modified beam search | 640ms      | 9.92       | 10.72      | --epoch 30 --avg 9  | chunk-wize           |
 
 Note: `simulated streaming` indicates feeding full utterance during decoding using `decode.py`,
 while `chunk-size` indicates feeding certain number of frames at each time using `streaming_decode.py`.
@@ -68,3 +68,49 @@ for m in greedy_search modified_beam_search fast_beam_search; do
     --num-decode-streams 2000
 done
 ```
+
+### zipformer (Zipformer + pruned statelss transducer)
+
+#### [zipformer](./zipformer)
+
+Number of model parameters: 74,778,511, i.e., 74.78 M
+
+##### Training on KsponSpeech (with MUSAN)
+
+Model: [johnBamma/icefall-asr-ksponspeech-zipformer-2024-06-24](https://huggingface.co/johnBamma/icefall-asr-ksponspeech-zipformer-2024-06-24)
+
+The CERs are:
+
+| decoding method      | eval_clean | eval_other | comment             |
+|----------------------|------------|------------|---------------------|
+| greedy search        | 10.60      | 11.56      | --epoch 30 --avg 9  |
+| fast beam search     | 10.59      | 11.54      | --epoch 30 --avg 9  |
+| modified beam search | 10.35      | 11.35      | --epoch 30 --avg 9  |
+
+The training command is:
+
+```bash
+./zipformer/train.py \
+    --world-size 4 \
+    --num-epochs 30 \
+    --start-epoch 1 \
+    --use-fp16 1 \
+    --exp-dir zipformer/exp \
+    --max-duration 750 \
+    --enable-musan True \
+    --base-lr 0.035
+```
+
+NOTICE: I decreased `base_lr` from 0.045(default) to 0.035, Because of `RuntimeError: grad_scale is too small`.
+
+The decoding command is:
+
+```bash
+for m in greedy_search fast_beam_search modified_beam_search; do
+    ./zipformer/decode.py \
+        --epoch 30 \
+        --avg 9 \
+        --exp-dir zipformer/exp \
+        --decoding-method $m
+done
+```

egs/ksponspeech/ASR/zipformer/README.md

@@ -0,0 +1 @@
+This recipe implements Zipformer model.

egs/ksponspeech/ASR/zipformer/asr_datamodule.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/asr_datamodule.py

egs/ksponspeech/ASR/zipformer/beam_search.py

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

egs/ksponspeech/ASR/zipformer/ctc_decode.py

@@ -0,0 +1,844 @@
+#!/usr/bin/env python3
+# Copyright    2024   (Author: SeungHyun Lee, Contacts: whsqkaak@naver.com)
+#
+# 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) ctc-decoding
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+
+(2) 1best
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method 1best
+
+(3) nbest
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method nbest
+
+(4) nbest-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method nbest-rescoring
+
+(5) whole-lattice-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method whole-lattice-rescoring
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import KsponSpeechAsrDataModule
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    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=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="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    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(
+        "--decoding-method",
+        type=str,
+        default="ctc-decoding",
+        help="""Decoding method.
+        Supported values are:
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (2) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (3) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+          rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+          the highest score is the decoding result.
+        - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+          n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+          is the decoding result.
+          you have trained an RNN LM using ./rnn_lm/train.py
+        - (6) nbest-oracle. Its WER is the lower bound of any n-best
+          rescoring method can achieve. Useful for debugging n-best
+          rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hlg-scale",
+        type=float,
+        default=0.6,
+        help="""The scale to be applied to `hlg.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - 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`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.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)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=ctc_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.decoding_method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.decoding_method in ["1best", "nbest"]:
+        if params.decoding_method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.decoding_method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.decoding_method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.decoding_method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.decoding_method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = 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.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring 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 = "?"
+
+    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,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+        )
+
+        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 % 100 == 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_cers = 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 CERs, 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:
+            cer = write_error_stats(f, f"{test_set_name}-{key}", results, compute_CER=True)
+            test_set_cers[key] = cer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_cers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_cers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    KsponSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "ctc-decoding",
+        "1best",
+        "nbest",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "nbest-oracle",
+    )
+    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 params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    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}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    if params.decoding_method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        HLG.scores *= params.hlg_scale
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.decoding_method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.decoding_method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_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 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()
+
+    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
+    
+    ksponspeech = KsponSpeechAsrDataModule(args)
+
+    eval_clean_cuts = ksponspeech.eval_clean_cuts()
+    eval_other_cuts = ksponspeech.eval_other_cuts()
+
+    eval_clean_dl = ksponspeech.test_dataloaders(eval_clean_cuts)
+    eval_other_dl = ksponspeech.test_dataloaders(eval_other_cuts)
+    
+    test_sets = ["eval_clean", "eval_other"]
+    test_dl = [eval_clean_dl, eval_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/ksponspeech/ASR/zipformer/decode_stream.py

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

egs/ksponspeech/ASR/zipformer/decoder.py

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

egs/ksponspeech/ASR/zipformer/encoder_interface.py

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

egs/ksponspeech/ASR/zipformer/export-onnx-ctc.py

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

egs/ksponspeech/ASR/zipformer/export-onnx-streaming-ctc.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-streaming-ctc.py

egs/ksponspeech/ASR/zipformer/export-onnx-streaming.py

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

egs/ksponspeech/ASR/zipformer/export-onnx.py

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

egs/ksponspeech/ASR/zipformer/export.py

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

egs/ksponspeech/ASR/zipformer/generate_averaged_model.py

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

egs/ksponspeech/ASR/zipformer/joiner.py

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

egs/ksponspeech/ASR/zipformer/model.py

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

egs/ksponspeech/ASR/zipformer/onnx_check.py

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

egs/ksponspeech/ASR/zipformer/onnx_decode.py

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

egs/ksponspeech/ASR/zipformer/onnx_pretrained-streaming-ctc.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained-streaming-ctc.py

egs/ksponspeech/ASR/zipformer/onnx_pretrained-streaming.py

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

egs/ksponspeech/ASR/zipformer/onnx_pretrained.py

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

egs/ksponspeech/ASR/zipformer/onnx_pretrained_ctc.py

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

egs/ksponspeech/ASR/zipformer/optim.py

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

egs/ksponspeech/ASR/zipformer/pretrained.py

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

egs/ksponspeech/ASR/zipformer/pretrained_ctc.py

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

egs/ksponspeech/ASR/zipformer/scaling.py

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

egs/ksponspeech/ASR/zipformer/scaling_converter.py

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

egs/ksponspeech/ASR/zipformer/streaming_beam_search.py

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

egs/ksponspeech/ASR/zipformer/streaming_decode.py

@@ -0,0 +1,856 @@
+#!/usr/bin/env python3
+# Copyright    2024   (Author: SeungHyun Lee, Contacts: whsqkaak@naver.com)
+#
+# 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:
+./zipformer/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --causal 1 \
+  --chunk-size 32 \
+  --left-context-frames 256 \
+  --exp-dir ./zipformer/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+from asr_datamodule import KsponSpeechAsrDataModule
+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 import Tensor, nn
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    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="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""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 get_init_states(
+    model: nn.Module,
+    batch_size: int = 1,
+    device: torch.device = torch.device("cpu"),
+) -> List[torch.Tensor]:
+    """
+    Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+    is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+    states[-2] is the cached left padding for ConvNeXt module,
+    of shape (batch_size, num_channels, left_pad, num_freqs)
+    states[-1] is processed_lens of shape (batch,), which records the number
+    of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+    """
+    states = model.encoder.get_init_states(batch_size, device)
+
+    embed_states = model.encoder_embed.get_init_states(batch_size, device)
+    states.append(embed_states)
+
+    processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+    states.append(processed_lens)
+
+    return states
+
+
+def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
+    """Stack list of zipformer states that correspond to separate utterances
+    into a single emformer state, so that it can be used as an input for
+    zipformer when those utterances are formed into a batch.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance. For element-n,
+        state_list[n] is a list of cached tensors of all encoder layers. For layer-i,
+        state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1,
+        cached_val2, cached_conv1, cached_conv2).
+        state_list[n][-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+        state_list[n][-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+    """
+    batch_size = len(state_list)
+    assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0])
+    tot_num_layers = (len(state_list[0]) - 2) // 6
+
+    batch_states = []
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key = torch.cat(
+            [state_list[i][layer_offset] for i in range(batch_size)], dim=1
+        )
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn = torch.cat(
+            [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1 = torch.cat(
+            [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2 = torch.cat(
+            [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1 = torch.cat(
+            [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2 = torch.cat(
+            [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0
+        )
+        batch_states += [
+            cached_key,
+            cached_nonlin_attn,
+            cached_val1,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        ]
+
+    cached_embed_left_pad = torch.cat(
+        [state_list[i][-2] for i in range(batch_size)], dim=0
+    )
+    batch_states.append(cached_embed_left_pad)
+
+    processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
+    batch_states.append(processed_lens)
+
+    return batch_states
+
+
+def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
+    """Unstack the zipformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Note:
+      It is the inverse of :func:`stack_states`.
+
+    Args:
+        batch_states: A list of cached tensors of all encoder layers. For layer-i,
+          states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
+          cached_conv1, cached_conv2).
+          state_list[-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+          states[-1] is processed_lens of shape (batch,), which records the number
+          of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Returns:
+        state_list: A list of list. Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance.
+    """
+    assert (len(batch_states) - 2) % 6 == 0, len(batch_states)
+    tot_num_layers = (len(batch_states) - 2) // 6
+
+    processed_lens = batch_states[-1]
+    batch_size = processed_lens.shape[0]
+
+    state_list = [[] for _ in range(batch_size)]
+
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1_list = batch_states[layer_offset + 2].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2_list = batch_states[layer_offset + 3].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1_list = batch_states[layer_offset + 4].chunk(
+            chunks=batch_size, dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2_list = batch_states[layer_offset + 5].chunk(
+            chunks=batch_size, dim=0
+        )
+        for i in range(batch_size):
+            state_list[i] += [
+                cached_key_list[i],
+                cached_nonlin_attn_list[i],
+                cached_val1_list[i],
+                cached_val2_list[i],
+                cached_conv1_list[i],
+                cached_conv2_list[i],
+            ]
+
+    cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(cached_embed_left_pad_list[i])
+
+    processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(processed_lens_list[i])
+
+    return state_list
+
+
+def streaming_forward(
+    features: Tensor,
+    feature_lens: Tensor,
+    model: nn.Module,
+    states: List[Tensor],
+    chunk_size: int,
+    left_context_len: int,
+) -> Tuple[Tensor, Tensor, List[Tensor]]:
+    """
+    Returns encoder outputs, output lengths, and updated states.
+    """
+    cached_embed_left_pad = states[-2]
+    (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        cached_left_pad=cached_embed_left_pad,
+    )
+    assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+
+    # processed_mask is used to mask out initial states
+    processed_mask = torch.arange(left_context_len, device=x.device).expand(
+        x.size(0), left_context_len
+    )
+    processed_lens = states[-1]  # (batch,)
+    # (batch, left_context_size)
+    processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+    # Update processed lengths
+    new_processed_lens = processed_lens + x_lens
+
+    # (batch, left_context_size + chunk_size)
+    src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+    encoder_states = states[:-2]
+    (
+        encoder_out,
+        encoder_out_lens,
+        new_encoder_states,
+    ) = model.encoder.streaming_forward(
+        x=x,
+        x_lens=x_lens,
+        states=encoder_states,
+        src_key_padding_mask=src_key_padding_mask,
+    )
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    new_states = new_encoder_states + [
+        new_cached_embed_left_pad,
+        new_processed_lens,
+    ]
+    return encoder_out, encoder_out_lens, new_states
+
+
+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
+    chunk_size = int(params.chunk_size)
+    left_context_len = int(params.left_context_frames)
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []  # Used in fast-beam-search
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(chunk_size * 2)
+        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)
+
+    # Make sure the length after encoder_embed is at least 1.
+    # The encoder_embed subsample features (T - 7) // 2
+    # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+    tail_length = chunk_size * 2 + 7 + 2 * 3
+    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)
+
+    encoder_out, encoder_out_lens, new_states = streaming_forward(
+        features=features,
+        feature_lens=feature_lens,
+        model=model,
+        states=states,
+        chunk_size=chunk_size,
+        left_context_len=left_context_len,
+    )
+
+    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 = torch.tensor(processed_lens, device=device)
+        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,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = 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 = 100
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = get_init_states(model=model, batch_size=1, 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
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=30)
+        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(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                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(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            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_cers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{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 CERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            cer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True, compute_CER=True,
+            )
+            test_set_cers[key] = cer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"cer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_cers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_cers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    KsponSpeechAsrDataModule.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}"
+
+    assert params.causal, params.causal
+    assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
+    assert (
+        "," not in params.left_context_frames
+    ), "left_context_frames should be one value in decoding."
+    params.suffix += f"-chunk-{params.chunk_size}"
+    params.suffix += f"-left-context-{params.left_context_frames}"
+
+    # 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}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_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)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    ksponspeech = KsponSpeechAsrDataModule(args)
+
+    eval_clean_cuts = ksponspeech.eval_clean_cuts()
+    eval_other_cuts = ksponspeech.eval_other_cuts()
+
+    test_sets = ["eval_clean", "eval_other"]
+    test_cuts = [eval_clean_cuts, eval_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/ksponspeech/ASR/zipformer/subsampling.py

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

egs/ksponspeech/ASR/zipformer/test_scaling.py

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

egs/ksponspeech/ASR/zipformer/test_subsampling.py

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

egs/ksponspeech/ASR/zipformer/zipformer.py

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