Pith. sign in

REVIEW 3 major objections 6 minor 39 references

Improving Practical Aspects of End-to-End Multi-Talker Speech Recognition for Online and Offline Scenarios

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding a continuous speech separation front-end to end-to-end multi-talker ASR improves accuracy in heavily overlapped speech, and segment-based serialization makes offline transcripts more readable.

desk verdict A solid engineering paper on multi-talker ASR with a genuinely new transcript-ordering scheme, but the headline claim that CSS separation drives the gains is not cleanly isolated from an architecture change. read the letter →

arxiv 2506.14204 v1 pith:AV3UB6QX submitted 2025-06-17 eess.AS cs.CLcs.SD

Aswin Shanmugam Subramanian , Amit Das , Naoyuki Kanda , Jinyu Li , Xiaofei Wang , Yifan Gong This is my paper · ORCID
classification eess.AScs.CLcs.SD
keywords multi-talkerspeechrecognitionserializedoutputtrainingcontinuousseparationstreamingASRofflinecascadedencodersegment-basedSOTLibriCSS
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether an explicit single-channel speech-separation front-end helps end-to-end multi-talker ASR in practical online and offline settings. The authors' central claim is that continuous speech separation (CSS) improves accuracy where overlap is heavy, with the streaming Conformer Transducer improving 8–9% relative at 30–40% overlap and the best offline model reaching 9.93% average speaker-agnostic word error rate (SAgWER) on LibriCSS. They also argue that a cascaded-encoder model can serve both streaming and offline modes with one set of parameters, and that segment-based serialized output training (segSOT) produces more readable multi-talker transcripts without post-processing. If these claims hold, explicit separation is a worthwhile addition to end-to-end systems, not a step backward.

What carries the argument

The load-bearing object is a two-channel Conformer encoder: the first $N$ of its $L$ layers are duplicated per channel, their outputs are summed, and the remaining $L-N$ layers are shared, so the parameter count matches a standard single-channel encoder. The two channels receive the outputs of a CSS front-end, a conformer-based speech-separation network with a 2.4-second window and 0.8-second hop that estimates two overlap-free signals per chunk. This mechanism makes the paper's main claim concrete: explicit separation happens upstream while the recognizer stays end-to-end. The secondary mechanism, segSOT, splits an utterance into segments governed by $\alpha$ (maximum speech-activity length) and $\beta$ (maximum pause length) and orders segments by start time, letting overlapping speakers take turns in the transcript.

What would settle it

Run both the CSS-equipped and non-CSS models on a test set containing genuine three-way overlap—for example, constructed triple mixtures from single-speaker utterances or recorded three-person conversations—and compare SAgWER. If the CSS model is no better than the non-CSS baseline in that condition, or if one speaker's words are systematically absent, the central claim that the front-end improves accuracy in highly overlapping speech is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that feeding a continuous speech separation (CSS) front-end—a sliding-window network that estimates two overlap-free signals from each 2.4-second chunk—into an end-to-end multi-talker ASR model improves accuracy precisely where overlap is heavy, without increasing the recognizer's parameter count. On the monaural LibriCSS test set, the streaming CT-tSOT system with the CSS encoder improves from 16.47% to 14.97% SAgWER at 40% overlap and from 14.25% to 13.07% at 30% overlap, while degrading only slightly on clean speech; the offline S2S-segSOT model with the CSS encoder achieves the best average SAgWER of 9.93%. The paper further shows that a cascaded encoder with 12 causal and 6 non-causal layers can switch between a 160ms streaming first pass and a 5s-look-ahead second pass, with the second pass improving average SAgWER from 11.38% to 10.22%. Finally, segSOT—which orders segments delimited by $\alpha$ and $\beta$ parameters—yields transcripts with natural turn-taking and better readability while preserving the monotonicity the CTC auxiliary loss favors.

Load-bearing premise

The entire system inherits the CSS front-end's two-speaker ceiling: each 2.4-second chunk is assumed to contain at most two overlapping speakers, so if three or more people talk at once the front-end cannot represent the mixture and the recognizer will miss or garble speech.

Editorial extensions

If this is right

  • In streaming recognition, adding a CSS encoder cuts SAgWER by 8–9% relative at 30–40% overlap, so explicit separation is a practical option for live multi-talker captioning.
  • A single cascaded-encoder model can cover both online and offline modes; its 5s-look-ahead second pass improves average SAgWER from 11.38% to 10.22%, reducing the need for separate systems.
  • segSOT transcription ordering makes offline multi-talker transcripts easier to read and preserves turn-taking, though it complicates RNN-T alignment and slightly hurts cascaded CT accuracy.
  • The best offline S2S-segSOT model with CSS encoder reaches 9.93% average SAgWER, and degrades by only about 1% relative on single-speaker LibriCSS, showing multi-talker training can retain single-talker capacity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the CSS front-end assumes exactly two speakers per chunk, the benefit is likely to vanish or invert on three-way overlap; a natural extension is a separation network with variable speaker count or a speaker-count gate upstream.
  • The paper's evaluation caps overlap at 40% on LibriCSS, so the claimed advantage in real meetings with occasional three-way cross-talk is unquantified; scoring on held-out real meeting recordings would test that gap.
  • The segSOT parameters $\alpha=5$s and $\beta=0.5$s control the readability/accuracy tradeoff, and other meeting domains may need different values; a sweep of these parameters on downstream summarization quality would clarify when segSOT helps beyond raw WER.
  • Since the CSS encoder helps most as overlap grows but slightly hurts clean audio, a front-end that switches on only when an overlap detector fires could preserve the gains and avoid the clean-speech cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This paper addresses practical aspects of end-to-end (E2E) multi-talker speech recognition for both streaming and offline scenarios. The authors propose three improvements: (1) integrating a continuous speech separation (CSS) front-end into SOT-based E2E models via a two-channel Conformer encoder, for both Conformer Transducer (CT) and sequence-to-sequence (S2S) systems; (2) a two-pass cascaded encoder model that shares a causal encoder for streaming and a non-causal encoder for offline decoding; and (3) segment-based SOT (segSOT) transcription ordering for offline model readability and turn-taking. Models are fine-tuned from 30k-hour single-speaker seeds on simulated mixtures and real meetings (AMI, ICSI, in-house), and evaluated on the monaural LibriCSS test set. The results show that the CSS front-end improves SAgWER by 8-9% relative in 30-40% overlap conditions for CT-tSOT, and that S2S-segSOT with CSS achieves the best average SAgWER of 9.93%. A cascaded CT-tSOT with CSS achieves 10.22% at 5s latency, demonstrating a latency-accuracy trade-off within a single model. The paper also provides detailed latency estimates and compares against a single-speaker seed model on 0% overlap conditions.

Significance. The paper's empirical results are useful and the experimental setup is careful in several respects: external evaluation on LibriCSS, real meeting data in training, matched parameter counts between single- and two-channel encoders, and consistent latency reporting. If the causal claim about CSS holds, the work would strengthen the case for explicit separation front-ends in E2E multi-talker ASR, which is a topic of current interest. The segSOT proposal addresses a genuine problem of SOT readability. However, the main result is weakened by a confounded comparison: the CSS condition changes both the input features and the encoder architecture, so the improvement cannot be uniquely attributed to separation. In addition, the paper asserts a comparison against cascaded systems and an advantage for segSOT that are not backed by in-paper baselines. These issues warrant revision but are addressable.

major comments (3)
  1. [§2.1, Table 1 rows 1-2 and 6-7] The CSS condition in Table 1 differs from the no-CSS condition in two ways at once: the input features (mixed audio vs. two CSS-estimated signals) and the encoder architecture (single 18-layer Conformer vs. two 18-layer channel-dependent branches summed at layer N). Since the abstract attributes the improvement to separation ("by separating overlapped speech"), a control experiment that feeds the two-branch encoder with the same mixed signal in both channels (or otherwise isolates the separation) is necessary to rule out that the gain comes from the extra input stream or the changed topology. Without such a control, the 8-9% relative gains at 30-40% overlap and the 4.2% relative average gain for S2S-segSOT cannot be uniquely attributed to CSS.
  2. [Abstract and §3.3 (first paragraph)] The claim of challenging the conventional wisdom of E2E versus cascaded systems is not supported by an in-paper cascaded baseline. The only evidence cited is a prior work [12] that showed CT-tSOT performing similarly to or better than a cascaded system under a different training setup. Because the current models are trained with real AMI/ICSI and in-house meeting data, the earlier comparison is not directly transferable. The authors should either add a cascaded system (CSS + single-talker ASR) to Table 1, or rephrase the abstract to avoid making an untested comparison against cascaded systems.
  3. [§2.3 and Table 1 (rows 6-7)] The segSOT advantage is not empirically demonstrated in the paper. Table 1 only contains S2S-segSOT results; there is no S2S-sSOT or S2S-uSOT baseline trained under the same conditions. The qualitative discussion of readability, turn-taking, and CTC is plausible, but the paper provides no measurement of these properties or a WER comparison to existing orderings. An ablation comparing segSOT with at least one prior ordering (e.g., sSOT) is needed to substantiate this advertised improvement.
minor comments (6)
  1. [§2.1] In Section 2.1, the variables L and N are not defined in the running text; please specify that L is the number of Conformer layers and N is the number of channel-dependent layers.
  2. [§3.3] In Section 3.3, the notation "S2S→segSOT: 7.02→7.05" is ambiguous; it should clearly state which model is the single-talker baseline and which is the multi-talker S2S-segSOT model.
  3. [Table 1, row 5] The segSOT hyperparameters for the cascaded CT-segSOT model in row 5 of Table 1 are not given; only the S2S-segSOT model reports (α, β) = (5, 0.5). Please report the values used or state that they are the same.
  4. [§3.2.3] In Section 3.2.3, the phrase "weights of 1 and 0.2 respectively" for the label smoothing and CTC losses is unclear, since label smoothing is typically a scalar within the CE loss. Please clarify the exact loss formula.
  5. [§3.3] In Section 3.3, the paper describes the CT-tSOT average improvement as "marginal" but does not provide per-session variance or significance; given the small absolute difference (11.52 vs 11.38), such information would help the reader judge the robustness.
  6. [§2.1] The two-speaker assumption of the CSS front-end (Section 2.1) is stated, but the paper does not discuss how the system degrades when more than two speakers overlap in a 2.4s chunk; a statement of this limitation in the conclusion would be appropriate, and a test on a three-speaker condition would strengthen the analysis.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: Table 1 SAgWER values are external measurements; the only caveats are a non-isolated comparison and minor same-group citations, neither of which makes the result equivalent to its inputs.

full rationale

The paper's central evidence (Table 1) compares SOT-based ASR models on the external LibriCSS test set and reports SAgWERs computed with NIST asclite; these are direct measurements, not quantities fitted from the model's own parameters. The CSS front-end is a separately trained component (Section 3.2.1) with a stated two-speaker assumption, and enabling/disabling this front-end while retraining the recognizer is a legitimate system-level comparison, even if it changes both input streams and encoder topology. The 'prediction' that CSS helps at 30-40% overlap is not definitionally tied to the evaluation. The main self-references ([12] for the tSOT baseline claim and [13] for the CSS front-end) share authors with this paper, but they support background context and component choice rather than the central accuracy numbers; no uniqueness theorem or fitted constant is imported to force the conclusion. The absence of a control that feeds identical mixed audio to both encoder branches is an experimental attribution limitation, not a circular reduction, because no equation or construction makes the reported improvement equal to the input features. Therefore no specific circular step can be exhibited, and the appropriate finding is no significant circularity, with a small deduction for reliance on same-group prior results in the framing.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central results rest on hand-chosen segSOT parameters (alpha, beta), a chosen simulation mix ratio, and the two-speaker-per-chunk assumption of the CSS front-end. No new physical entities or forces are introduced; the evaluation uses the standard SAgWER metric on the external LibriCSS test set, which keeps the empirical claims independent of the training data.

free parameters (3)
  • alpha (max speech-activity segment length) = 5 s
    Hand-chosen for segSOT; controls when the transcript switches speakers and determines training labels. No sweep or sensitivity analysis is reported.
  • beta (max short-pause length) = 0.5 s
    Hand-chosen for segSOT; a pause longer than beta closes the current segment. No sweep or sensitivity analysis is reported.
  • Simulated overlap mixture ratio = 2/3 mixed, 1/3 single
    Hand-chosen training data proportion between overlapped and single-speaker utterances; affects the model's exposure to overlap during fine-tuning.
assumptions (3)
  • domain assumption Each CSS chunk contains at most two active speakers, so two estimated streams suffice.
    Section 2.1: 'assuming that there are only two speakers in the chunk'. If three or more speakers overlap within a chunk, the front-end cannot represent the mixture and the ASR will miss content; the paper reports no condition that violates or tests this.
  • domain assumption LibriCSS sessions 1-9, excluding session 0, are a fair held-out evaluation set.
    Section 3.1 states sessions 1 through 9 are used. The paper does not state whether session 0 was used for development or hyperparameter selection, so the evaluation set may not be fully independent.
  • domain assumption Simulated two-speaker mixing (about two-thirds of utterances) is representative of real overlap conditions.
    Section 3.1 describes the simulation recipe; the claim that the models generalize to LibriCSS rests on this representativeness alongside real AMI/ICSI data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Practical Aspects of End-to-End Multi-Talker Speech Recognition for Online and Offline Scenarios." pith.science (2026). https://pith.science/paper/AV3UB6QX

@misc{pith2026250614204,
  author       = {Pith},
  title        = {Pith review of: Improving Practical Aspects of End-to-End Multi-Talker Speech Recognition for Online and Offline Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AV3UB6QX}},
  note         = {Machine review of arXiv:2506.14204}
}
read the original abstract

We extend the frameworks of Serialized Output Training (SOT) to address practical needs of both streaming and offline automatic speech recognition (ASR) applications. Our approach focuses on balancing latency and accuracy, catering to real-time captioning and summarization requirements. We propose several key improvements: (1) Leveraging Continuous Speech Separation (CSS) single-channel front-end with end-to-end (E2E) systems for highly overlapping scenarios, challenging the conventional wisdom of E2E versus cascaded setups. The CSS framework improves the accuracy of the ASR system by separating overlapped speech from multiple speakers. (2) Implementing dual models -- Conformer Transducer for streaming and Sequence-to-Sequence for offline -- or alternatively, a two-pass model based on cascaded encoders. (3) Exploring segment-based SOT (segSOT) which is better suited for offline scenarios while also enhancing readability of multi-talker transcriptions.

Figures

Figures reproduced from arXiv: 2506.14204 by the authors.

Figure 2
Figure 2. Conformer Transducer with Multi-Talker Cascaded Encoder. The Causal Encoder here takes two channel CSS in￾puts. the entire input sequence x, it is suitable for modeling offline scenarios. There are different ways of ordering/serializing the tran￾scriptions in multi-talker simulations. One such way is the SOT paradigm [10,24]. Transcriptions of a multi-talker conversation are shown in [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 3
Figure 3. Multi-talker Transcription (α = Maximum length of speech activity, β = Maximum length of short pause). lays in transcribing other overlapping speakers thereby allowing frequent turn-taking scenarios which is common in multi-talker conversations. Following this, speaker 2 pauses for a while that exceeds β seconds. Thus, segSOT stops transcribing speaker 2 and switches to the earliest available segment which is from s… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 35 canonical work pages

  1. [12]

    Streaming multi-talker ASR with token-level serialized output training,

    N. Kanda, J. Wu, Y . Wu, X. Xiao, Z. Meng, X. Wang, Y . Gaur, Z. Chen, J. Li, and T. Yoshioka, “Streaming multi-talker ASR with token-level serialized output training,” inProc. Interspeech, 2022, pp. 3774–3778

    work page 2022

  2. [1]

    However, the challenge of recognizing overlapping speech in multi-talker sce- narios remains a critical area of research

    Introduction End-to-end (E2E) automatic speech recognition (ASR) [1, 2] has made significant strides in recent years, achieving remark- able performance on various benchmarks [3–5]. However, the challenge of recognizing overlapping speech in multi-talker sce- narios remains a critical area of research. Traditional ASR systems struggle with overlapping spe...

    work page

  3. [2]

    Improving Practical Aspects of End-to-End Multi-Talker Speech Recognition for Online and Offline Scenarios

    Model 2.1. Conformer Encoder with CSS Input We used an encoder architecture similar to what was used with a multichannel front-end in [16]. As shown in Fig. 1, the en- coder is designed to process two audio signals by splitting the conformer encoder withLlayers in total intoNchannel de- arXiv:2506.14204v1 [eess.AS] 17 Jun 2025 pendent layers andL−Nchannel...

    work page Pith review arXiv 2025

  4. [3]

    Avg. ” column. 0L and 0S are 0% overlap conditions with long and short inter-utterance silences. Column “CSS

    Experiments 3.1. Data We trained two seed single-speaker ASR models (CT, S2S) us- ing 30,000 hours of in-house data [27], with all personally iden- tifiable information removed. To develop multi-speaker mod- els, we fine-tuned the initial seed model using a diverse dataset. This dataset included: a) simulated multi-speaker data derived from the aforementi...

    work page

  5. [4]

    First, we leverage speech separated signals by using two channel CSS en- coder in our ASR models

    Conclusion In this paper, we have proposed three improvements in E2E multi-talker ASR modeling that effectively balances latency and accuracy for both streaming and offline scenarios. First, we leverage speech separated signals by using two channel CSS en- coder in our ASR models. Second, we implemented online and offline models and introduced a unified t...

    work page

  6. [5]

    Clearly, this model significantly outperforms the CT-tSOT model in row 2 across all scenarios

    Next, in row 4, we compare the second pass results of the cascaded CT-tSOT model using both the causal and non-causal encoder layers. Clearly, this model significantly outperforms the CT-tSOT model in row 2 across all scenarios. The improve- ment is 10.2% relative (11.38→10.2) at the cost of more la- tency. This demonstrates a trade-off between accuracy a...

    work page

  7. [6]

    Continuous speech separation: dataset and analysis,

    Z. Chen, T. Yoshioka, L. Lu, T. Zhou, Z. Meng, Y . Luo, J. Wu, X. Xiao, and J. Li, “Continuous speech separation: dataset and analysis,” inProc. ICASSP, 2020, pp. 7284–7288

    work page 2020

  8. [7]

    Recent advances in end-to-end automatic speech recogni- tion,

    J. Li, “Recent advances in end-to-end automatic speech recogni- tion,”APSIPA Transactions on Signal and Information Process- ing, vol. 11, no. 1, 2022

    work page 2022

Show all 39 references
  1. [8]

    End-to-end speech recognition: A survey,

    R. Prabhavalkar, T. Hori, T. N. Sainath, R. Schl¨uter, and S. Watan- abe, “End-to-end speech recognition: A survey,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 325–351, 2023

    2023

  2. [9]

    A streaming on-device end-to-end model surpassing server-side conventional model quality and latency,

    T. N. Sainath, Y . He, B. Li, A. Narayanan, R. Pang, A. Bruguier, S.-y. Chang, W. Li, R. Alvarez, Z. Chenet al., “A streaming on-device end-to-end model surpassing server-side conventional model quality and latency,” inProc. ICASSP. IEEE, 2020, pp. 6059–6063

    2020

  3. [10]

    Developing RNN- T models surpassing high-performance hybrid models with cus- tomization capability,

    J. Li, R. Zhao, Z. Meng, Y . Liu, W. Wei, S. Parthasarathy, V . Mazalov, Z. Wang, L. He, S. Zhaoet al., “Developing RNN- T models surpassing high-performance hybrid models with cus- tomization capability,” inProc. Interspeech, 2020, pp. 3590– 3594

    2020

  4. [11]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” OpenAI, Tech. Rep., 2022

    2022

  5. [13]

    Integration of speech separation, diarization, and recognition for multi-speaker meetings: System description, comparison, and analysis,

    D. Raj, P. Denisov, Z. Chen, H. Erdogan, Z. Huang, M. He, S. Watanabe, J. Du, T. Yoshioka, Y . Luoet al., “Integration of speech separation, diarization, and recognition for multi-speaker meetings: System description, comparison, and analysis,” inProc. SLT, 2021, pp. 897–904

    2021

  6. [14]

    Permutation invari- ant training of deep models for speaker-independent multi-talker speech separation,

    D. Yu, M. Kolbæk, Z. Tan, and J. Jensen, “Permutation invari- ant training of deep models for speaker-independent multi-talker speech separation,” inIEEE International Conference on Acous- tics, Speech, and Signal Processing (ICASSP), 2017, pp. 241–245

    2017

  7. [15]

    Recognizing multi-talker speech with permutation invariant training,

    D. Yu, X. Chang, and Y . Qian, “Recognizing multi-talker speech with permutation invariant training,”Proc. Interspeech 2017, pp. 2456–2460, 2017

    2017

  8. [16]

    Figure 2:Conformer Transducer with Multi-Talker Cascaded Encoder

    utilized multi-channel CSS. Figure 2:Conformer Transducer with Multi-Talker Cascaded Encoder. The Causal Encoder here takes two channel CSS in- puts. the entire input sequencex, it is suitable for modeling offline scenarios. There are different ways of ordering/serializing the...

  9. [17]

    Seri- alized output training for end-to-end overlapped speech recogni- tion,

    N. Kanda, Y . Gaur, X. Wang, Z. Meng, and T. Yoshioka, “Seri- alized output training for end-to-end overlapped speech recogni- tion,” inProc. Interspeech, 2020, pp. 2797–2801

    2020

  10. [18]

    Rec- ognizing overlapped speech in meetings: A multichannel separa- tion approach using neural networks,

    T. Yoshioka, H. Erdogan, Z. Chen, X. Xiao, and F. Alleva, “Rec- ognizing overlapped speech in meetings: A multichannel separa- tion approach using neural networks,” inProc. Interspeech, 2018, p. 3038–3042

    2018

  11. [19]

    Speech separation with large-scale self-supervised learning,

    Z. Chen, N. Kanda, J. Wu, Y . Wu, X. Wang, T. Yoshioka, J. Li, S. Sivasankaran, and S. E. Eskimez, “Speech separation with large-scale self-supervised learning,” inICASSP 2023-2023 IEEE international conference on acoustics, speech and signal process- ing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  12. [20]

    MIMO-Speech: End-to-end multi-channel multi-speaker speech recognition,

    X. Chang, W. Zhang, Y . Qian, J. Le Roux, and S. Watanabe, “MIMO-Speech: End-to-end multi-channel multi-speaker speech recognition,” inIEEE Automatic Speech Recognition and Under- standing Workshop (ASRU), 2019, pp. 237–244

    2019

  13. [21]

    Directional ASR: A new paradigm for E2E multi-speaker speech recognition with source localization,

    A. S. Subramanian, C. Weng, S. Watanabe, M. Yu, Y . Xu, S.- X. Zhang, and D. Yu, “Directional ASR: A new paradigm for E2E multi-speaker speech recognition with source localization,” inIEEE International Conference on Acoustics, Speech, and Sig- nal Processing (ICASSP), 2021, p...

    2021

  14. [22]

    VarArray meets t-SOT: Advancing the state of the art of stream- ing distant conversational speech recognition,

    N. Kanda, J. Wu, X. Wang, Z. Chen, J. Li, and T. Yoshioka, “VarArray meets t-SOT: Advancing the state of the art of stream- ing distant conversational speech recognition,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  15. [23]

    End-to-end multi-speaker speech recognition with transformer,

    X. Chang, W. Zhang, Y . Qian, J. L. Roux, and S. Watanabe, “End-to-end multi-speaker speech recognition with transformer,” inProceedings of the IEEE International Conference on Acous- tics, Speech, and Signal Processing (ICASSP), 2020, pp. 6134– 6138

    2020

  16. [24]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented transformer for speech recognition,” in Proceedings of Interspeech. ISCA, 2020, pp. 5036–5040

    2020

  17. [25]

    Developing real-time streaming transformer transducer for speech recognition on large- scale dataset,

    X. Chen, Y . Wu, Z. Wang, S. Liu, and J. Li, “Developing real-time streaming transformer transducer for speech recognition on large- scale dataset,” inProc. ICASSP. IEEE, 2021, pp. 5904–5908

    2021

  18. [26]

    End-to- end continuous speech recognition using attention-based recurrent NN: First results,

    J. Chorowski, D. Bahdanau, K. Cho, and Y . Bengio, “End-to- end continuous speech recognition using attention-based recurrent NN: First results,”arXiv preprint arXiv:1412.1602, 2014

    2014 arXiv

  19. [27]

    Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,

    W. Chan, N. Jaitly, Q. Le, and O. Vinyals, “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” inProc. ICASSP, 2016, pp. 4960–4964

    2016

  20. [28]

    Cascaded encoders for unifying streaming and non-streaming ASR,

    A. Narayanan, T. N. Sainath, R. Pang, J. Yu, C.-C. Chiu, R. Prab- havalkar, E. Variani, and T. Strohman, “Cascaded encoders for unifying streaming and non-streaming ASR,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 5629–5633

    2021

  21. [29]

    Continuous speech separation with con- former,

    S. Chen, Y . Wu, Z. Chen, J. Wu, J. Li, T. Yoshioka, C. Wang, S. Liu, and M. Zhou, “Continuous speech separation with con- former,” inProc. ICASSP, 2021, pp. 5749–5753

    2021

  22. [30]

    Joint speaker counting, speech recognition, and speaker identification for overlapped speech of any number of speakers,

    N. Kanda, Y . Gaur, X. Wang, Z. Meng, Z. Chen, T. Zhou, and T. Yoshioka, “Joint speaker counting, speech recognition, and speaker identification for overlapped speech of any number of speakers,” inProc. Interspeech, 2020, pp. 36–40

    2020

  23. [31]

    Investigation of end-to-end speaker-attributed ASR for continuous multi-talker recordings,

    N. Kanda, X. Chang, Y . Gaur, X. Wang, Z. Meng, Z. Chen, and T. Yoshioka, “Investigation of end-to-end speaker-attributed ASR for continuous multi-talker recordings,” inIEEE SLT, 2021, pp. 809–816

    2021

  24. [32]

    Joint CTC-attention based end-to-end speech recognition using multi-task learning,

    S. Kim, T. Hori, and S. Watanabe, “Joint CTC-attention based end-to-end speech recognition using multi-task learning,” in ICASSP. IEEE, 2017, pp. 4835–4839

    2017

  25. [33]

    High-accuracy and low-latency speech recognition with two-head contextual layer trajectory LSTM model,

    J. Li, R. Zhao, E. Sun, J. H. Wong, A. Das, Z. Meng, and Y . Gong, “High-accuracy and low-latency speech recognition with two-head contextual layer trajectory LSTM model,” inICASSP. IEEE, 2020, pp. 7699–7703

    2020

  26. [34]

    The AMI meeting corpus: A pre- announcement,

    J. Carletta, S. Ashby, and et al., “The AMI meeting corpus: A pre- announcement,” inInternational Workshop on Machine Learning for Multimodal Interaction, 2006, pp. 28–39

    2006

  27. [35]

    The ICSI meeting corpus,

    A. Janin, D. Baron, and et al., “The ICSI meeting corpus,” inProc. ICASSP, 2003, pp. 364–367

    2003

  28. [36]

    The NIST Scoring Toolkit (SCTK),

    “The NIST Scoring Toolkit (SCTK),” 2021. [Online]. Available: https://www.nist.gov/itl/iad/mig/tools

    2021

  29. [37]

    WavLM: Large-scale self- supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “WavLM: Large-scale self- supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  30. [38]

    Handling trade-offs in speech separation with sparsely-gated mixture of experts,

    X. Wang, Z. Chen, Y . Shi, J. Wu, N. Kanda, and T. Yoshioka, “Handling trade-offs in speech separation with sparsely-gated mixture of experts,” inarXiv preprint arXiv:2211.06493, 2022

    2022 arXiv

  31. [39]

    Improving wideband speech recognition using mixed-bandwidth training data in CD- DNN-HMM,

    J. Li, D. Yu, J.-T. Huang, and Y . Gong, “Improving wideband speech recognition using mixed-bandwidth training data in CD- DNN-HMM,” inProceedings of SLT. IEEE, 2012, pp. 131–136

    2012

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.