zipformer/ctc_align.py

- tool for forced-alignment with CTC model
- provides timeline, computes per-token and per-utterance acoustic confidences
- based on torchaudio `forced_align()`
- confidences are computed in several ways

other modifications:
- LibriSpeechAsrDataModel extended with `::load_manifest()` to allow
  passing-in cutset from CLI.
- update @custom_fwd @custom_bwd in scaling.py
- streaming_decode.py update errs/recogs/log filenames '-' <-> '_'

egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py

@@ -402,6 +402,14 @@ class LibriSpeechAsrDataModule:
         )
         return test_dl
 
+    @lru_cache()
+    def load_manifest(self, manifest_filename: str) -> CutSet:
+        """
+        Load the 'manifest' specified by an argument.
+        """
+        logging.info(f"About to get '{manifest_filename}' cuts")
+        return load_manifest_lazy(manifest_filename)
+
     @lru_cache()
     def train_clean_5_cuts(self) -> CutSet:
         logging.info("mini_librispeech: About to get train-clean-5 cuts")

egs/librispeech/ASR/zipformer/ctc_align.py

@@ -0,0 +1,661 @@
+#!/usr/bin/env python3
+#
+# Copyright 2025  Brno University of Technology (Author: Karel Vesely)
+#
+# 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.
+
+"""
+Batch aligning with CTC model (it can be Tranducer + CTC).
+It works with both causal an non-causal models.
+Streaming is disabled, or simulated by attention masks
+(see: --chunk-size --left-context-frames).
+Whole utterance processed by 1 forward() call.
+
+Note: model averaging is present. With `--epoch 10 --avg 3`,
+the epochs 8-10 are taken for averaging. Model averaging
+is smoothing the CTC posteriors to some extent.
+
+Usage:
+(1) torchaudio forced_align()
+./zipformer/ctc_align.py \
+    --epoch 10 \
+    --avg 3 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 300 \
+    --decoding-method ctc_align
+
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path, PurePath
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule as AsrDataModule
+from lhotse import set_caching_enabled
+from torchaudio.functional import (
+    forced_align,
+    merge_tokens,
+)
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    str2bool,
+)
+
+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(
+        "--res-dir-suffix",
+        type=str,
+        default="",
+        help="Suffix to where alignments are stored",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--ignored-tokens",
+        type=str,
+        nargs="+",
+        default=[],
+        help="",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="ctc_align",
+        choices=[
+            "ctc_align",
+        ],
+        help=""" Decoding method for doing the forced alignment.""",
+    )
+
+    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(
+        "dataset_manifests",
+        type=str,
+        nargs="+",
+        help="""Manifests of test-sets to be evaluated""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def align_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    ignored_tokens: set[int],
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Align one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for alignment.
+               For now, just "ctc_alignment" is used.
+        - value: It contains the alignment result: (labels, log_probs).
+                 `len(value)` equals to batch size. `value[i]` is the alignment
+                 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.
+      ignored_tokens:
+        Set of int token-codes to be ignored for calculation of confidence.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+
+      UNUSED_PART, CAN BE USED LATER FOR ALIGNING TO A DECODING_GRAPH:
+
+      word_table [UNUSED]:
+        The word symbol table.
+      decoding_graph [UNUSED]:
+        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 alignment 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)
+
+    batch_size = feature.shape[0]
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    hyps = []
+
+    # tokenize the transcripts:
+    text_encoded = sp.encode(supervisions["text"])
+
+    # lengths
+    num_tokens = [len(te) for te in text_encoded]
+    max_tokens = max(num_tokens)
+
+    # convert to padded np.array:
+    targets = np.array(
+        [
+            np.pad(seq, (0, max_tokens - len(seq)), "constant", constant_values=-1)
+            for seq in text_encoded
+        ]
+    )
+
+    # convert to tensor:
+    targets = torch.tensor(targets, dtype=torch.int32, device=device)
+    target_lengths = torch.tensor(num_tokens, dtype=torch.int32, device=device)
+
+    # torchaudio2.4.0+
+    # The batch dimension for log_probs must be 1 at the current version:
+    # https://github.com/pytorch/audio/blob/main/src/libtorchaudio/forced_align/gpu/compute.cu#L277
+    for ii in range(batch_size):
+        labels, log_probs = forced_align(
+            log_probs=ctc_output[ii, : encoder_out_lens[ii]].unsqueeze(dim=0),
+            targets=targets[ii, : target_lengths[ii]].unsqueeze(dim=0),
+            input_lengths=encoder_out_lens[ii].unsqueeze(dim=0),
+            target_lengths=target_lengths[ii].unsqueeze(dim=0),
+            blank=0,
+        )
+
+        # per-token time, score
+        token_spans = merge_tokens(labels[0], log_probs[0].exp())
+        # int -> token
+        for s in token_spans:
+            s.token = sp.id_to_piece(s.token)
+        # mean conf. from the per-token scores
+        mean_token_conf = np.mean([token_span.score for token_span in token_spans])
+
+        # confidences
+        ignore_mask = labels == 0
+        for tok in ignored_tokens:
+            ignore_mask += labels == tok
+
+        nonblank_scores = log_probs[~ignore_mask].exp()
+        num_scores = nonblank_scores.shape[0]
+
+        if num_scores > 0:
+            nonblank_min = float(nonblank_scores.min())
+            nonblank_q05 = float(torch.quantile(nonblank_scores, 0.05))
+            nonblank_q10 = float(torch.quantile(nonblank_scores, 0.10))
+            nonblank_q20 = float(torch.quantile(nonblank_scores, 0.20))
+            nonblank_q30 = float(torch.quantile(nonblank_scores, 0.30))
+            nonblank_mean = float(nonblank_scores.mean())
+        else:
+            nonblank_min = -1.0
+            nonblank_q05 = -1.0
+            nonblank_q10 = -1.0
+            nonblank_q20 = -1.0
+            nonblank_q30 = -1.0
+            nonblank_mean = -1.0
+
+        if num_scores > 0:
+            confidence = (nonblank_min + nonblank_q05 + nonblank_q10 + nonblank_q20) / 4
+        else:
+            confidence = 1.0  # default score for short utts
+
+        hyps.append(
+            {
+                "token_spans": token_spans,
+                "mean_token_conf": mean_token_conf,
+                "confidence": confidence,
+                "num_scores": num_scores,
+                "nonblank_mean": nonblank_mean,
+                "nonblank_min": nonblank_min,
+                "nonblank_q05": nonblank_q05,
+                "nonblank_q10": nonblank_q10,
+                "nonblank_q20": nonblank_q20,
+                "nonblank_q30": nonblank_q30,
+            }
+        )
+
+    return {"ctc_align": hyps}
+
+
+def align_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 = "?"
+
+    ignored_tokens = params.ignored_tokens + ["<sos/eos>", "<unk>"]
+    ignored_tokens_ints = [sp.piece_to_id(token) for token in ignored_tokens]
+
+    logging.info(f"ignored tokens {ignored_tokens}")
+    logging.info(f"ignored int codes {ignored_tokens_ints}")
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = align_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            ignored_tokens=ignored_tokens_ints,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, alignments in hyps_dict.items():
+            this_batch = []
+            assert len(alignments) == len(texts)
+            for cut_id, alignments, ref_text in zip(cut_ids, alignments, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, alignments))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        log_interval = 100
+        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_alignment_output(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    """
+    Save the token alignments and per-utterance confidences.
+    """
+
+    for key, results in results_dict.items():
+
+        alignments_filename = params.res_dir / f"alignments-{test_set_name}.txt"
+
+        time_step = 0.04
+
+        with open(alignments_filename, "w", encoding="utf8") as fd:
+            for key, ref_text, ali in results:
+                for token_span in ali["token_spans"]:
+
+                    t_beg = token_span.start * time_step
+                    t_end = token_span.end * time_step
+                    t_dur = t_end - t_beg
+                    token = token_span.token
+                    score = token_span.score
+
+                    # CTM format : (wav_name, ch, t_beg, t_dur, token, score)
+                    print(
+                        f"{key} A {t_beg:.2f} {t_dur:.2f} {token} {score:.6f}", file=fd
+                    )
+
+        logging.info(f"The alignments are stored in `{alignments_filename}`")
+
+        # ---------------------------
+
+        confidences_filename = params.res_dir / f"confidences-{test_set_name}.txt"
+
+        with open(confidences_filename, "w", encoding="utf8") as fd:
+            print(
+                "utterance_key mean_token_conf mean_frame_conf q0-20_conf "
+                "(nonblank_min,q05,q10,q20,q30) (num_scores,num_tokens)",
+                file=fd,
+            )  # header
+            for key, ref_text, ali in results:
+                mean_token_conf = ali["mean_token_conf"]
+                mean_frame_conf = ali["nonblank_mean"]
+                q0_20_conf = ali["confidence"]
+                min_ = ali["nonblank_min"]
+                q05 = ali["nonblank_q05"]
+                q10 = ali["nonblank_q10"]
+                q20 = ali["nonblank_q20"]
+                q30 = ali["nonblank_q30"]
+                num_scores = ali[
+                    "num_scores"
+                ]  # scores used to compute `mean_frame_conf`
+                num_tokens = len(ali["token_spans"])  # tokens in ref transcript
+                print(
+                    f"{key} {mean_token_conf:.4f} {mean_frame_conf:.4f} "
+                    f"{q0_20_conf:.4f} "
+                    f"({min_:.4f},{q05:.4f},{q10:.4f},{q20:.4f},{q30:.4f}) "
+                    f"({num_scores},{num_tokens})",
+                    file=fd,
+                )
+
+        logging.info(f"The confidences are stored in `{confidences_filename}`")
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    # enable AudioCache
+    set_caching_enabled(True)  # lhotse
+
+    assert params.decoding_method in ("ctc_align",)
+    assert params.enable_spec_aug is False
+    assert params.use_ctc is True
+
+    params.res_dir = params.exp_dir / (params.decoding_method + params.res_dir_suffix)
+
+    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}"
+
+    params.suffix += f"_{params.decoding_method}"
+
+    if params.use_averaged_model:
+        params.suffix += "_use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-align-{params.suffix}")
+    logging.info("Forced-alignment 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()
+
+    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
+    asr_datamodule = AsrDataModule(args)
+
+    # create array of dataloaders (one per test-set)
+    testset_labels = []
+    testset_dataloaders = []
+    for testset_manifest in args.dataset_manifests:
+        label = PurePath(testset_manifest).name  # basename
+        label = label.replace(".jsonl.gz", "")
+
+        test_cuts = asr_datamodule.load_manifest(testset_manifest)
+        test_dataloader = asr_datamodule.test_dataloaders(test_cuts)
+
+        testset_labels.append(label)
+        testset_dataloaders.append(test_dataloader)
+
+    # align
+    for test_set, test_dl in zip(testset_labels, testset_dataloaders):
+        results_dict = align_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=None,
+            decoding_graph=None,
+        )
+
+        save_alignment_output(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/zipformer/scaling.py

@@ -24,7 +24,7 @@ import k2
 import torch
 import torch.nn as nn
 from torch import Tensor
-from torch.cuda.amp import custom_bwd, custom_fwd
+from torch.amp import custom_bwd, custom_fwd
 
 from icefall.utils import torch_autocast
 
@@ -1306,7 +1306,7 @@ class MulForDropout3(torch.autograd.Function):
     # returns (x * y * alpha) where alpha is a float and y doesn't require
     # grad and is zero-or-one.
     @staticmethod
-    @custom_fwd
+    @custom_fwd(device_type='cuda')
     def forward(ctx, x, y, alpha):
         assert not y.requires_grad
         ans = x * y * alpha
@@ -1315,7 +1315,7 @@ class MulForDropout3(torch.autograd.Function):
         return ans
 
     @staticmethod
-    @custom_bwd
+    @custom_bwd(device_type='cuda')
     def backward(ctx, ans_grad):
         (ans,) = ctx.saved_tensors
         x_grad = ctx.alpha * ans_grad * (ans != 0)
@@ -1512,7 +1512,7 @@ def SwooshRForward(x: Tensor):
 
 class ActivationDropoutAndLinearFunction(torch.autograd.Function):
     @staticmethod
-    @custom_fwd
+    @custom_fwd(device_type='cuda')
     def forward(
         ctx,
         x: Tensor,
@@ -1551,7 +1551,7 @@ class ActivationDropoutAndLinearFunction(torch.autograd.Function):
         return x
 
     @staticmethod
-    @custom_bwd
+    @custom_bwd(device_type='cuda')
     def backward(ctx, ans_grad: Tensor):
         saved = ctx.saved_tensors
         (x, weight, bias, dropout_mask) = saved

egs/librispeech/ASR/zipformer/streaming_decode.py

@@ -641,15 +641,15 @@ def decode_dataset(
             del decode_streams[i]
 
     if params.decoding_method == "greedy_search":
-        key = "greedy_search"
+        key = "greedy-search"
     elif params.decoding_method == "fast_beam_search":
         key = (
-            f"beam_{params.beam}_"
-            f"max_contexts_{params.max_contexts}_"
-            f"max_states_{params.max_states}"
+            f"beam-{params.beam}_"
+            f"max-contexts-{params.max_contexts}_"
+            f"max-states-{params.max_states}"
         )
     elif params.decoding_method == "modified_beam_search":
-        key = f"num_active_paths_{params.num_active_paths}"
+        key = f"num-active-paths-{params.num_active_paths}"
     else:
         raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
     return {key: decode_results}
@@ -665,7 +665,7 @@ def save_asr_output(
     """
     for key, results in results_dict.items():
         recogs_filename = (
-            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+            params.res_dir / f"recogs-{test_set_name}_{key}_{params.suffix}.txt"
         )
         results = sorted(results)
         store_transcripts(filename=recogs_filename, texts=results)
@@ -685,11 +685,11 @@ def save_wer_results(
         # 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"
+            params.res_dir / f"errs-{test_set_name}_{key}_{params.suffix}.txt"
         )
         with open(errs_filename, "w", encoding="utf8") as fd:
             wer = write_error_stats(
-                fd, f"{test_set_name}-{key}", results, enable_log=True
+                fd, f"{test_set_name}_{key}", results, enable_log=True
             )
             test_set_wers[key] = wer
 
@@ -698,7 +698,7 @@ def save_wer_results(
     test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
 
     wer_filename = (
-        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+        params.res_dir / f"wer-summary_{test_set_name}_{key}_{params.suffix}.txt"
     )
     with open(wer_filename, "w", encoding="utf8") as fd:
         print("settings\tWER", file=fd)
@@ -729,9 +729,9 @@ def main():
     params.res_dir = params.exp_dir / "streaming" / params.decoding_method
 
     if params.iter > 0:
-        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+        params.suffix = f"iter-{params.iter}_avg-{params.avg}"
     else:
-        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+        params.suffix = f"epoch-{params.epoch}_avg-{params.avg}"
 
     assert params.causal, params.causal
     assert "," not in params.chunk_size, "chunk_size should be one value in decoding."