Merge cc168d104128348e9e24835c856c1bd946638e71 into 231bbcd2b638826a94cf019fa31ae8683d3552ee

egs/libriheavy/ASR/local/bpe2tokens.py

@@ -0,0 +1,37 @@
+#!/usr/bin/env python3
+
+"""
+This script takes `bpe.model` as input and generates a file `tokens.txt`
+from it.
+
+Usage:
+./bpe_model_to_tokens.py /path/to/input/bpe.model > tokens.txt
+"""
+import argparse
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "bpe_model",
+        type=str,
+        help="Path to the input bpe.model",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    for i in range(sp.vocab_size()):
+        print(sp.id_to_piece(i), i)
+
+
+if __name__ == "__main__":
+    main()

egs/libriheavy/ASR/local/compute_fbank_libriheavy.py

@@ -0,0 +1,241 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# 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.
+
+
+"""
+This file computes fbank features of the LibriSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    LilcomChunkyWriter,
+)
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--fbank-dir",
+        type=str,
+        help="""Fbank output dir
+        """,
+        default="data/fbank",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the medium and large subset.",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Process pieces starting from this number (inclusive).",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop processing pieces until this number (exclusive).",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_libriheavy(
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = ("small",)
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for part in dataset_parts:
+            output_cuts_path = output_dir / f"librilight_cuts_{part}.jsonl.gz"
+            if output_cuts_path.exists():
+                logging.info(f"{output_cuts_path} exists - skipping")
+                continue
+
+            input_cuts_path = src_dir / f"librilight_cuts_{part}.jsonl.gz"
+            assert input_cuts_path.exists(), f"{input_cuts_path} does not exist!"
+            logging.info(f"Loading {input_cuts_path}")
+            cut_set = CutSet.from_file(input_cuts_path)
+
+            logging.info("Computing features")
+
+            if bpe_model:
+                cut_set = filter_cuts(cut_set, sp)
+
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/libriheavy_feats_{part}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+
+            logging.info(f"Saving to {output_cuts_path}")
+            cut_set.to_file(output_cuts_path)
+
+
+def compute_fbank_libriheavy_splits(args):
+    num_splits = args.num_splits
+    dataset = args.dataset
+    output_dir = f"{args.fbank_dir}/libriheavy_{dataset}_split"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = len(str(num_splits))
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    # if torch.cuda.is_available():
+    #     device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    prefix = "libriheavy"
+
+    num_digits = 8  # num_digits is fixed by lhotse split-lazy
+    for i in range(start, stop):
+        idx = f"{i + 1}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"{prefix}_cuts_{dataset}.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"{prefix}_cuts_{dataset}_raw.{idx}.jsonl.gz"
+        if not raw_cuts_path.is_file():
+            logging.info(f"{raw_cuts_path} does not exist - skipping it")
+            continue
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Computing features")
+        if (output_dir / f"{prefix}_feats_{dataset}_{idx}.lca").exists():
+            logging.info(f"Removing {output_dir}/{prefix}_feats_{dataset}_{idx}.lca")
+            os.remove(output_dir / f"{prefix}_feats_{dataset}_{idx}.lca")
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_{dataset}_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            overwrite=True,
+        )
+
+        logging.info("About to split cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+        logging.info(f"Saved to {cuts_path}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+
+    compute_fbank_libriheavy_splits(args)

egs/libriheavy/ASR/local/filter_cuts.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/filter_cuts.py

egs/libriheavy/ASR/local/prepare_lang.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py

egs/libriheavy/ASR/local/prepare_lang_bpe.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py

egs/libriheavy/ASR/local/prepare_validation_sets.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# 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.
+
+
+"""
+This file computes fbank features of the LibriSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+from lhotse import load_manifest_lazy
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--in-manifest", type=str, help="The original manifest coming from"
+    )
+
+    parser.add_argument(
+        "--out-manifest",
+        type=str,
+        help="Where to store the manifest after filtering out the test/dev sets",
+    )
+
+    return parser.parse_args()
+
+
+def main(args):
+
+    logging.info(f"Loading manifest {args.in_manifest}")
+    cuts = load_manifest_lazy(args.in_manifest)
+
+    all_test_sets = [
+        "dev",
+        "test-clean",
+        "test-other",
+    ]
+
+    all_books = []
+    for test_set in all_test_sets:
+        logging.info(f"Processing test set: {test_set}")
+        with open(f"data/manifests/{test_set}.txt", "r") as f:
+            books = f.read().split("\n")
+        all_books += books
+
+        out_name = f"data/manifests/libriheavy_cuts_{test_set}.jsonl.gz"
+        if os.path.exists(out_name):
+            continue
+        # find the cuts belonging to the given books
+        selected_cuts = cuts.filter(lambda c: c.text_path.split("/")[-2] in books)
+        selected_cuts.describe()
+
+        logging.info(f"Saving the cuts contained in the book list to {out_name}")
+        selected_cuts.to_file(out_name)
+
+    filtered_cuts = cuts.filter(lambda c: c.text_path.split("/")[-2] not in all_books)
+    logging.info(f"Saving the filtered manifest to {args.out_manifest}.")
+    filtered_cuts.to_file(args.out_manifest)
+    logging.info("Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+
+    main(args)

egs/libriheavy/ASR/local/train_bpe_model.py

@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+from icefall.utils import str2bool
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    parser.add_argument(
+        "--byte-fallback",
+        type=str2bool,
+        default=False,
+    )
+
+    parser.add_argument(
+        "--character-coverage",
+        type=float,
+        default=0.99,
+        help="Character coverage when training BPE",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    print(args)
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    character_coverage = 1.0
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=args.character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+            train_extremely_large_corpus=False,
+            byte_fallback=args.byte_fallback,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()

egs/libriheavy/ASR/prepare.sh

@@ -0,0 +1,166 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+export PYTHONPATH=/star-data/xiaoyu/icefall_libriheavy:$PYTHONPATH
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+start=0
+stop=-1
+num_per_split=2000
+split_per_job=20
+char_coverage=0.99
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  750
+)
+
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+manifest_dir=data/manifests
+fbank_dir=data/fbank
+
+mkdir -p $manifest_dir
+
+subset="medium"
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Split libri-heavy ${subset}"
+
+  if [ $subset == "large" ]; then
+    num_per_split=8000
+    log "Change num_per_split to ${num_per_split} for large"
+  fi
+
+  split_dir=$fbank_dir/libriheavy_${subset}_split
+  mkdir -p $split_dir
+  if [ ! -e $split_dir/.split_completed ]; then
+    lhotse split-lazy $manifest_dir/libriheavy_cuts_${subset}_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.split_completed
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compute fbank for Libri-heavy ${subset}"
+  mkdir -p $fbank_dir
+  num_splits=$(find $fbank_dir/libriheavy_${subset}_split -name "libriheavy_cuts_${subset}_raw.*.jsonl.gz" | wc -l)
+  if [ $subset == "large" ]; then
+    split_per_job=210
+    log "Change split_per_job to ${split_per_job} for large"
+  elif [ $subset == "medium" ]; then
+    split_per_job=100
+    log "Change split_per_job to ${split_per_job} for medium"
+  fi
+  if [ ! -e $fbank_dir/.libriheavy.${subset}.done ]; then
+    for i in $(seq 0 1 7); do
+      start=$(( i * $split_per_job ))
+      end=$(( (i+1) * $split_per_job ))
+      ./local/compute_fbank_libriheavy.py \
+        --dataset ${subset} \
+        --fbank-dir $fbank_dir \
+        --num-splits $num_splits \
+        --num-workers $nj \
+        --start $start \
+        --stop $end &
+    done
+    wait
+    touch $fbank_dir/.libriheavy.${subset}.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Combine features for ${subset}"
+  if [ ! -f $fbank_dir/libriheavy_cuts_${subset}.jsonl.gz ]; then
+    pieces=$(find $fbank_dir/libriheavy_${subset}_split -name "libriheavy_cuts_${subset}.*.jsonl.gz")
+    lhotse combine $pieces $fbank_dir/libriheavy_cuts_${subset}.jsonl.gz
+  fi
+fi
+
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Prepare the validation&test sets"
+
+  ./local/prepare_validation_sets.py \
+    --in-manifest $fbank_dir/libriheavy_cuts_medium.jsonl.gz \
+    --out-manifest $fbank_dir/libriheavy_cuts_medium_filtered.jsonl.gz
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare BPE model"
+
+  tmp_dir=data/tmp
+  mkdir -p $tmp_dir
+  if [ ! -f $tmp_dir/transcript_words.txt ]; then
+    for part in "small" "medium" "large"; do
+      gunzip -c $manifest_dir/libriheavy_cuts_${part}_raw.jsonl.gz |
+        jq '.supervisions[].custom.texts[]' | sed 's/" //' | sed 's/\(.*\)"/\1/' > $tmp_dir/transcript_words_${part}.txt
+    done
+    cat $tmp_dir/transcript_words_small.txt $tmp_dir/transcript_words_medium.txt $tmp_dir/transcript_words_large.txt > $tmp_dir/transcript_words.txt
+  fi
+
+  if [ ! -f $tmp_dir/words.txt ]; then
+    cat $tmp_dir/transcript_words.txt | sed 's/ /\n/g' \
+      | sort -u | sed '/^$/d' > $tmp_dir/words.txt
+    (echo '!SIL'; echo '<SPOKEN_NOISE>'; echo '<UNK>'; ) |
+    cat - $tmp_dir/words.txt | sort | uniq | awk '
+    BEGIN {
+      print "<eps> 0";
+    }
+    {
+      if ($1 == "<s>") {
+
+        print "<s> is in the vocabulary!" | "cat 1>&2"
+        exit 1;
+      }
+      if ($1 == "</s>") {
+        print "</s> is in the vocabulary!" | "cat 1>&2"
+        exit 1;
+      }
+      printf("%s %d\n", $1, NR);
+    }
+    END {
+      printf("#0 %d\n", NR+1);
+      printf("<s> %d\n", NR+2);
+      printf("</s> %d\n", NR+3);
+    }' > $tmp_dir/words || exit 1;
+    mv $tmp_dir/words $tmp_dir/words.txt
+  fi
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}_fallback_coverage_${char_coverage}
+    mkdir -p $lang_dir
+    cp $tmp_dir/words.txt $lang_dir/words.txt
+    pushd $lang_dir
+    ln -s ../$tmp_dir/transcript_words.txt transcript_words.txt
+    popd
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --byte-fallback True \
+        --character-coverage $char_coverage \
+        --transcript $tmp_dir/transcript_words_medium.txt
+    fi
+
+    if [ ! -f $lang_dir/tokens.txt ]; then
+      ./local/bpe2tokens.py ${lang_dir}/bpe.model > ${lang_dir}/tokens.txt
+    fi
+
+    done
+fi

egs/libriheavy/ASR/zipformer/asr_datamodule.py

@@ -0,0 +1,585 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    load_manifest,
+    load_manifest_lazy,
+    validate,
+)
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    SingleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (
+    BatchIO,
+    OnTheFlyFeatures,
+    PrecomputedFeatures,
+)
+from lhotse.utils import fix_random_seed, ifnone
+from text_normalization import replace_full_width_symbol, simple_normalization
+from torch.utils.data.dataloader import DataLoader, default_collate
+
+from icefall.utils import str2bool
+
+
+class LibriHeavyASRDataset(torch.utils.data.Dataset):
+    """This is a dataset for LibriHeavy dataset"""
+
+    def __init__(
+        self,
+        return_cuts: bool = False,
+        cut_transforms: List[Callable[[CutSet], CutSet]] = None,
+        input_transforms: List[Callable[[torch.Tensor], torch.Tensor]] = None,
+        input_strategy: BatchIO = PrecomputedFeatures(),
+        text_sampling_func: Optional[Callable[[List[str]], str]] = None,
+    ):
+        """
+        Icefall ASR IterableDataset constructor. See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py
+        for more details.
+
+        :param return_cuts: When ``True``, will additionally return a "cut" field in each batch with the Cut
+            objects used to create that batch.
+        :param cut_transforms: A list of transforms to be applied on each sampled batch,
+            before converting cuts to an input representation (audio/features).
+            Examples: cut concatenation, noise cuts mixing, etc.
+        :param input_transforms: A list of transforms to be applied on each sampled batch,
+            after the cuts are converted to audio/features.
+            Examples: normalization, SpecAugment, etc.
+        :param input_strategy: Converts cuts into a collated batch of audio/features.
+            By default, reads pre-computed features from disk.
+        :param text_sampling_func: Sampling a text as transcription from a list of texts.
+        """
+        super().__init__()
+        # Initialize the fields
+        self.return_cuts = return_cuts
+        self.cut_transforms = ifnone(cut_transforms, [])
+        self.input_transforms = ifnone(input_transforms, [])
+        self.input_strategy = input_strategy
+
+        # a text selection function
+        self.text_sampling_func = text_sampling_func
+
+    def __getitem__(self, cuts: CutSet) -> Dict[str, Union[torch.Tensor, List[str]]]:
+        """
+        Return a new batch, with the batch size automatically determined using the constraints
+        of max_frames and max_cuts.
+        """
+        validate_for_asr(cuts)
+
+        # Sort the cuts by duration so that the first one determines the batch time dimensions.
+        cuts = cuts.sort_by_duration(ascending=False)
+
+        # Optional CutSet transforms - e.g. padding, or speed perturbation that adjusts
+        # the supervision boundaries.
+        for tnfm in self.cut_transforms:
+            cuts = tnfm(cuts)
+
+        # Sort the cuts again after transforms
+        cuts = cuts.sort_by_duration(ascending=False)
+
+        # Get a tensor with batched feature matrices, shape (B, T, F)
+        # Collation performs auto-padding, if necessary.
+        input_tpl = self.input_strategy(cuts)
+        if len(input_tpl) == 3:
+            # An input strategy with fault tolerant audio reading mode.
+            # "cuts" may be a subset of the original "cuts" variable,
+            # that only has cuts for which we succesfully read the audio.
+            inputs, _, cuts = input_tpl
+        else:
+            inputs, _ = input_tpl
+
+        # Get a dict of tensors that encode the positional information about supervisions
+        # in the batch of feature matrices. The tensors are named "sequence_idx",
+        # "start_frame/sample" and "num_frames/samples".
+        supervision_intervals = self.input_strategy.supervision_intervals(cuts)
+
+        # Apply all available transforms on the inputs, i.e. either audio or features.
+        # This could be feature extraction, global MVN, SpecAugment, etc.
+        segments = torch.stack(list(supervision_intervals.values()), dim=1)
+        for tnfm in self.input_transforms:
+            inputs = tnfm(inputs, supervision_segments=segments)
+
+        batch = {
+            "inputs": inputs,
+            "supervisions": default_collate(
+                [
+                    simple_normalization(
+                        self.text_sampling_func(texts=supervision.texts)
+                    )
+                    if self.text_sampling_func is not None
+                    else {
+                        "text": simple_normalization(supervision.texts[0]),
+                    }
+                    for sequence_idx, cut in enumerate(cuts)
+                    for supervision in cut.supervisions
+                ]
+            ),
+        }
+        # Update the 'supervisions' field with sequence_idx and start/num frames/samples
+        batch["supervisions"].update(supervision_intervals)
+        if self.return_cuts:
+            batch["supervisions"]["cut"] = [
+                cut for cut in cuts for sup in cut.supervisions
+            ]
+
+        has_word_alignments = all(
+            s.alignment is not None and "word" in s.alignment
+            for c in cuts
+            for s in c.supervisions
+        )
+
+        return batch
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriHeavyAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it "
+            "with training dataset. ",
+        )
+
+        # Libriheavy specific arguments
+        group.add_argument(
+            "--subset",
+            type=str,
+            default="small",
+            help="Select the Libriheavy subset (small|medium|large)",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+        text_sampling_func: Optional[Callable[[List[str]], str]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            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)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = LibriHeavyASRDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+            text_sampling_func=text_sampling_func,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = LibriHeavyASRDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+                text_sampling_func=text_sampling_func,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SingleCutSampler.")
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(
+        self,
+        cuts_valid: CutSet,
+        text_sampling_func: Optional[Callable[[List[str]], str]] = None,
+    ) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = LibriHeavyASRDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+                text_sampling_func=text_sampling_func,
+            )
+        else:
+            validate = LibriHeavyASRDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+                text_sampling_func=text_sampling_func,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = LibriHeavyASRDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info(f"About to get {self.args.subset} cuts")
+
+        path = self.args.manifest_dir / "libriheavy_cuts_small.jsonl.gz"
+        cuts_train = CutSet.from_jsonl_lazy(path)
+        if self.args.subset == "medium":
+            logging.info("Getting medium subset")
+            path = self.args.manifest_dir / "libriheavy_cuts_medium.jsonl.gz"
+            cuts_train += CutSet.from_jsonl_lazy(path)
+        elif self.args.subset == "large":
+            logging.info("Getting large subset")
+            path = self.args.manifest_dir / "libriheavy_cuts_medium.jsonl.gz"
+            cuts_train += CutSet.from_jsonl_lazy(path)
+            path = self.args.manifest_dir / "libriheavy_cuts_large.jsonl.gz"
+            cuts_train += CutSet.from_jsonl_lazy(path)
+
+        return cuts_train
+
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_dev.jsonl.gz"
+        )
+        return cuts
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test-clean.jsonl.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test-other.jsonl.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def librispeech_test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def librispeech_test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz"
+        )
+
+
+def validate_for_asr(cuts: CutSet) -> None:
+    validate(cuts)
+    tol = 2e-3  # 1ms
+    for cut in cuts:
+        for supervision in cut.supervisions:
+            assert supervision.start >= -tol, (
+                f"Supervisions starting before the cut are not supported for ASR"
+                f" (sup id: {supervision.id}, cut id: {cut.id})"
+            )
+
+            # Supervision start time is relative to Cut ...
+            # https://lhotse.readthedocs.io/en/v0.10_e/cuts.html
+            #
+            # 'supervision.end' is end of supervision inside the Cut
+            assert supervision.end <= cut.duration + tol, (
+                f"Supervisions ending after the cut "
+                f"are not supported for ASR"
+                f" (sup id: {supervision.id}, cut id: {cut.id})"
+            )

egs/libriheavy/ASR/zipformer/beam_search.py

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

egs/libriheavy/ASR/zipformer/decode.py

@@ -0,0 +1,875 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# 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
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --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)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --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)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+import warnings
+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 LibriHeavyAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from text_normalization import (
+    simple_normalization,
+    upper_normalization,
+    word_normalization,
+)
+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.lexicon import Lexicon
+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=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(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    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=20.0,
+        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,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=False,
+        help="""Upper case and remove all chars except ' and -
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    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)
+
+    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)
+
+    hyps = []
+
+    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,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            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,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            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,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            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,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: 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.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    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)
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        for batch_idx, batch in enumerate(dl):
+            texts = batch["supervisions"]["text"]
+            texts = [
+                simple_normalization(t) for t in texts
+            ]  # Do a simple normalization, as this is done during training
+
+            cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+            hyps_dict = decode_one_batch(
+                params=params,
+                model=model,
+                sp=sp,
+                decoding_graph=decoding_graph,
+                word_table=word_table,
+                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}-{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 WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    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 "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}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    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}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are 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 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 "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = 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
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    def add_texts(c: Cut):
+        text = c.supervisions[0].text
+        c.supervisions[0].texts = [text]
+        return c
+
+    test_clean_cuts = libriheavy.test_clean_cuts()
+    test_other_cuts = libriheavy.test_other_cuts()
+    ls_test_clean_cuts = libriheavy.librispeech_test_clean_cuts()
+    ls_test_other_cuts = libriheavy.librispeech_test_other_cuts()
+
+    ls_test_clean_cuts = ls_test_clean_cuts.map(add_texts)
+    ls_test_other_cuts = ls_test_other_cuts.map(add_texts)
+
+    test_clean_dl = libriheavy.test_dataloaders(test_clean_cuts)
+    test_other_dl = libriheavy.test_dataloaders(test_other_cuts)
+    ls_test_clean_dl = libriheavy.test_dataloaders(ls_test_clean_cuts)
+    ls_test_other_dl = libriheavy.test_dataloaders(ls_test_other_cuts)
+
+    # test_sets = ["libriheavy-test-clean", "libriheavy-test-other", "librispeech-test-clean", "librispeech-test-other"]
+    # test_dl = [test_clean_dl, test_other_dl, ls_test_clean_dl, ls_test_other_dl]
+
+    test_sets = ["librispeech-test-clean", "librispeech-test-other"]
+    test_dl = [ls_test_clean_dl, ls_test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+        if params.post_normalization:
+            params.suffix += "-post-normalization"
+
+            new_res = {}
+            for k in results_dict:
+                new_ans = []
+                for item in results_dict[k]:
+                    id, ref, hyp = item
+                    hyp = upper_normalization(" ".join(hyp)).split()
+                    hyp = [word_normalization(w) for w in hyp]
+                    hyp = " ".join(hyp).split()
+                    hyp = [w for w in hyp if w != ""]
+                    new_ans.append((id, ref, hyp))
+                new_res[k] = new_ans
+
+            save_results(
+                params=params,
+                test_set_name=test_set,
+                results_dict=new_res,
+            )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/libriheavy/ASR/zipformer/decoder.py

@@ -0,0 +1,123 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from scaling import Balancer
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+        )
+        # the balancers are to avoid any drift in the magnitude of the
+        # embeddings, which would interact badly with parameter averaging.
+        self.balancer = Balancer(decoder_dim, channel_dim=-1,
+                                 min_positive=0.0, max_positive=1.0,
+                                 min_abs=0.5, max_abs=1.0,
+                                 prob=0.05)
+
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=decoder_dim,
+                out_channels=decoder_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=decoder_dim // 4,  # group size == 4
+                bias=False,
+            )
+            self.balancer2 = Balancer(decoder_dim, channel_dim=-1,
+                                      min_positive=0.0, max_positive=1.0,
+                                      min_abs=0.5, max_abs=1.0,
+                                      prob=0.05)
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        y = y.to(torch.int64)
+        # this stuff about clamp() is a temporary fix for a mismatch
+        # at utterance start, we use negative ids in beam_search.py
+        embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
+
+        embedding_out = self.balancer(embedding_out)
+
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(
+                    embedding_out, pad=(self.context_size - 1, 0)
+                )
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+            embedding_out = F.relu(embedding_out)
+            embedding_out = self.balancer2(embedding_out)
+
+        return embedding_out

egs/libriheavy/ASR/zipformer/encoder_interface.py

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

egs/libriheavy/ASR/zipformer/joiner.py

@@ -0,0 +1,66 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from scaling import ScaledLinear
+
+
+class Joiner(nn.Module):
+    def __init__(
+        self,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        super().__init__()
+
+        self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim, initial_scale=0.25)
+        self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim, initial_scale=0.25)
+        self.output_linear = nn.Linear(joiner_dim, vocab_size)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        project_input: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+           project_input:
+            If true, apply input projections encoder_proj and decoder_proj.
+            If this is false, it is the user's responsibility to do this
+            manually.
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim, (encoder_out.shape, decoder_out.shape)
+
+        if project_input:
+            logit = self.encoder_proj(encoder_out) + self.decoder_proj(
+                decoder_out
+            )
+        else:
+            logit = encoder_out + decoder_out
+
+        logit = self.output_linear(torch.tanh(logit))
+
+        return logit

egs/libriheavy/ASR/zipformer/model.py

@@ -0,0 +1,358 @@
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       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.
+
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos, make_pad_mask
+from scaling import ScaledLinear
+
+
+class AsrModel(nn.Module):
+    def __init__(
+        self,
+        encoder_embed: nn.Module,
+        encoder: EncoderInterface,
+        decoder: Optional[nn.Module] = None,
+        joiner: Optional[nn.Module] = None,
+        encoder_dim: int = 384,
+        decoder_dim: int = 512,
+        vocab_size: int = 500,
+        use_transducer: bool = True,
+        use_ctc: bool = False,
+    ):
+        """A joint CTC & Transducer ASR model.
+
+        - Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks (http://imagine.enpc.fr/~obozinsg/teaching/mva_gm/papers/ctc.pdf)
+        - Sequence Transduction with Recurrent Neural Networks (https://arxiv.org/pdf/1211.3711.pdf)
+        - Pruned RNN-T for fast, memory-efficient ASR training (https://arxiv.org/pdf/2206.13236.pdf)
+
+        Args:
+          encoder_embed:
+            It is a Convolutional 2D subsampling module. It converts
+            an input of shape (N, T, idim) to an output of of shape
+            (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2.
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dim) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+            It is used when use_transducer is True.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+            It is used when use_transducer is True.
+          use_transducer:
+            Whether use transducer head. Default: True.
+          use_ctc:
+            Whether use CTC head. Default: False.
+        """
+        super().__init__()
+
+        assert (
+            use_transducer or use_ctc
+        ), f"At least one of them should be True, but got use_transducer={use_transducer}, use_ctc={use_ctc}"
+
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+
+        self.encoder_embed = encoder_embed
+        self.encoder = encoder
+
+        self.use_transducer = use_transducer
+        if use_transducer:
+            # Modules for Transducer head
+            assert decoder is not None
+            assert hasattr(decoder, "blank_id")
+            assert joiner is not None
+
+            self.decoder = decoder
+            self.joiner = joiner
+
+            self.simple_am_proj = ScaledLinear(
+                encoder_dim, vocab_size, initial_scale=0.25
+            )
+            self.simple_lm_proj = ScaledLinear(
+                decoder_dim, vocab_size, initial_scale=0.25
+            )
+        else:
+            assert decoder is None
+            assert joiner is None
+
+        self.use_ctc = use_ctc
+        if use_ctc:
+            # Modules for CTC head
+            self.ctc_output = nn.Sequential(
+                nn.Dropout(p=0.1),
+                nn.Linear(encoder_dim, vocab_size),
+                nn.LogSoftmax(dim=-1),
+            )
+
+    def forward_encoder(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute encoder outputs.
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+
+        Returns:
+          encoder_out:
+            Encoder output, of shape (N, T, C).
+          encoder_out_lens:
+            Encoder output lengths, of shape (N,).
+        """
+        # logging.info(f"Memory allocated at entry: {torch.cuda.memory_allocated() // 1000000}M")
+        x, x_lens = self.encoder_embed(x, x_lens)
+        # logging.info(f"Memory allocated after encoder_embed: {torch.cuda.memory_allocated() // 1000000}M")
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        assert torch.all(encoder_out_lens > 0), (x_lens, encoder_out_lens)
+
+        return encoder_out, encoder_out_lens
+
+    def forward_ctc(
+        self,
+        encoder_out: torch.Tensor,
+        encoder_out_lens: torch.Tensor,
+        targets: torch.Tensor,
+        target_lengths: torch.Tensor,
+    ) -> torch.Tensor:
+        """Compute CTC loss.
+        Args:
+          encoder_out:
+            Encoder output, of shape (N, T, C).
+          encoder_out_lens:
+            Encoder output lengths, of shape (N,).
+          targets:
+            Target Tensor of shape (sum(target_lengths)). The targets are assumed
+            to be un-padded and concatenated within 1 dimension.
+        """
+        # Compute CTC log-prob
+        ctc_output = self.ctc_output(encoder_out)  # (N, T, C)
+
+        ctc_loss = torch.nn.functional.ctc_loss(
+            log_probs=ctc_output.permute(1, 0, 2),  # (T, N, C)
+            targets=targets,
+            input_lengths=encoder_out_lens,
+            target_lengths=target_lengths,
+            reduction="sum",
+        )
+        return ctc_loss
+
+    def forward_transducer(
+        self,
+        encoder_out: torch.Tensor,
+        encoder_out_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        y_lens: torch.Tensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute Transducer loss.
+        Args:
+          encoder_out:
+            Encoder output, of shape (N, T, C).
+          encoder_out_lens:
+            Encoder output lengths, of shape (N,).
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        """
+        # Now for the decoder, i.e., the prediction network
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (encoder_out.size(0), 4),
+            dtype=torch.int64,
+            device=encoder_out.device,
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = encoder_out_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        # if self.training and random.random() < 0.25:
+        #    lm = penalize_abs_values_gt(lm, 100.0, 1.0e-04)
+        # if self.training and random.random() < 0.25:
+        #    am = penalize_abs_values_gt(am, 30.0, 1.0e-04)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return simple_loss, pruned_loss
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer losses and CTC loss,
+          in form of (simple_loss, pruned_loss, ctc_loss)
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        # Compute encoder outputs
+        encoder_out, encoder_out_lens = self.forward_encoder(x, x_lens)
+
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        if self.use_transducer:
+            # Compute transducer loss
+            simple_loss, pruned_loss = self.forward_transducer(
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                y=y.to(x.device),
+                y_lens=y_lens,
+                prune_range=prune_range,
+                am_scale=am_scale,
+                lm_scale=lm_scale,
+            )
+        else:
+            simple_loss = torch.empty(0)
+            pruned_loss = torch.empty(0)
+
+        if self.use_ctc:
+            # Compute CTC loss
+            targets = y.values
+            ctc_loss = self.forward_ctc(
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                targets=targets,
+                target_lengths=y_lens,
+            )
+        else:
+            ctc_loss = torch.empty(0)
+
+        return simple_loss, pruned_loss, ctc_loss

egs/libriheavy/ASR/zipformer/optim.py

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

egs/libriheavy/ASR/zipformer/scaling.py

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

egs/libriheavy/ASR/zipformer/scaling_converter.py

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

egs/libriheavy/ASR/zipformer/subsampling.py

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

egs/libriheavy/ASR/zipformer/text_normalization.py

@@ -0,0 +1,148 @@
+import re
+
+words = {
+    0: "zero",
+    1: "one",
+    2: "two",
+    3: "three",
+    4: "four",
+    5: "five",
+    6: "six",
+    7: "seven",
+    8: "eight",
+    9: "nine",
+    10: "ten",
+    11: "eleven",
+    12: "twelve",
+    13: "thirteen",
+    14: "fourteen",
+    15: "fifteen",
+    16: "sixteen",
+    17: "seventeen",
+    18: "eighteen",
+    19: "nineteen",
+    20: "twenty",
+    30: "thirty",
+    40: "forty",
+    50: "fifty",
+    60: "sixty",
+    70: "seventy",
+    80: "eighty",
+    90: "ninety",
+}
+ordinal_nums = [
+    "zeroth",
+    "first",
+    "second",
+    "third",
+    "fourth",
+    "fifth",
+    "sixth",
+    "seventh",
+    "eighth",
+    "ninth",
+    "tenth",
+    "eleventh",
+    "twelfth",
+    "thirteenth",
+    "fourteenth",
+    "fifteenth",
+    "sixteenth",
+    "seventeenth",
+    "eighteenth",
+    "nineteenth",
+    "twentieth",
+]
+
+num_ordinal_dict = {num: ordinal_nums[num] for num in range(21)}
+
+
+def year_to_words(num: int):
+    assert isinstance(num, int), num
+    # check if a num is representing a year
+    if num > 1500 and num < 2000:
+        return words[num // 100] + " " + num_to_words(num % 100)
+    elif num == 2000:
+        return "TWO THOUSAND"
+    elif num > 2000:
+        return "TWO THOUSAND AND " + num_to_words(num % 100)
+    else:
+        return num_to_words(num)
+
+
+def num_to_words(num: int):
+    # Return the English words of a integer number
+
+    # If this is a year number
+    if num > 1500 and num < 2030:
+        return year_to_words(num)
+
+    if num < 20:
+        return words[num]
+    if num < 100:
+        if num % 10 == 0:
+            return words[num // 10 * 10]
+        else:
+            return words[num // 10 * 10] + " " + words[num % 10]
+    if num < 1000:
+        return words[num // 100] + " hundred and " + num_to_words(num % 100)
+    if num < 1000000:
+        return num_to_words(num // 1000) + " thousand " + num_to_words(num % 1000)
+    return num
+
+
+def num_to_ordinal_word(num: int):
+
+    return num_ordinal_dict.get(num, num_to_words(num)).upper()
+
+
+def replace_full_width_symbol(s: str) -> str:
+    # replace full-width symbol with theri half width counterpart
+    s = s.replace("“", '"')
+    s = s.replace("”", '"')
+    s = s.replace("‘", "'")
+    s = s.replace("’", "'")
+
+    return s
+
+
+def upper_normalization(text: str) -> str:
+    text = replace_full_width_symbol(text)
+    text = text.upper()  # upper case all characters
+
+    # Only keep all alpha-numeric characters, hypen and apostrophe
+    text = text.replace("-", " ")
+    text = re.sub("[^a-zA-Z0-9\s']+", "", text)
+    return text
+
+
+def word_normalization(word: str) -> str:
+    if word == "MRS":
+        return "MISSUS"
+    if word == "MR":
+        return "MISTER"
+    if word == "ST":
+        return "SAINT"
+    if word == "ECT":
+        return "ET CETERA"
+    if word.isnumeric():
+        word = num_to_words(int(word))
+        return word.upper()
+    if word[-2:] == "TH" and word[0].isnumeric():  #  e.g 9TH, 6TH
+        return num_to_ordinal_word(int(word[:-2])).upper()
+
+    return word
+
+
+def simple_normalization(text: str) -> str:
+    text = replace_full_width_symbol(text)
+    text = text.replace("--", " ")
+
+    return text
+
+
+if __name__ == "__main__":
+
+    s = str(1830)
+    out = word_normalization(s)
+    print(s, out)

egs/libriheavy/ASR/zipformer/zipformer.py

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