Pith. sign in

REVIEW 2 major objections 5 minor 41 references

The paper argues that, because speech firmly anchors the transcript, an LLM decoder can safely emit multiple tokens per step, turning ASR's decoding bottleneck into a verification regime.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 10:07 UTC pith:HYOXRNEV

load-bearing objection Useful ASR systems idea, but the headline acceptance metric is miscalculated under the paper's own rejection rule and the ablation is confounded; needs revision before it can be trusted. the 2 major comments →

arxiv 2607.29279 v1 pith:HYOXRNEV submitted 2026-07-31 cs.SD

ParaASR: Multi-Token Prediction for Fast and Long-Context LLM-Based Speech Recognition

Qingjian Lin , Yuxin Li , Haoyang Zhang , Jun Chen , Yechang Huang , Feng Tian , Xie Li , Xiangyu Tony Zhang
show 9 more authors
Daijiao Liu Yuxin Zhang Jinglan Gong Bo Zhao Fei Tian Xuerui Yang Gang Yu Xiangyu Zhang Daxin Jiang
This is my paper
classification cs.SD
keywords automatic speech recognitionmulti-token predictionLLM-based ASRspeculative decodinglong-form transcriptionreal-time factordecoding accelerationaudio-language model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that the usual trade-off between decoder scale and decoding speed in LLM-based speech recognition is avoidable. Because a transcript is anchored to the audio signal, the next several tokens are highly predictable, so a decoder can emit many tokens per forward step rather than one. To exploit this, the authors build ParaASR: a 4-billion-parameter decoder with five auxiliary multi-token prediction branches, trained after the main recognizer converges. At inference the model proposes six tokens per step and admits only the verified prefix, so the transcript stays identical to standard autoregressive decoding. The paper reports an average of 5.0 accepted tokens per step, a real-time factor of 0.0053, and native 32K-context transcription of up to 30 minutes, claiming decoder scaling, low latency, and long-context capability are not competing goals.

Core claim

The central claim is that ASR is a naturally parallel decoding problem because the output is firmly conditioned on acoustics. The paper implements this via multi-token prediction: five branches attached to the decoder's last layer forecast tokens at offsets 1 through 5 alongside the main next-token head. After the autoregressive recognizer is fully trained, the branches are aligned with the frozen backbone and then jointly fine-tuned, with losses weighted by an exponential decay. During serving, all six proposed tokens are checked against the autoregressive distribution and the longest matching prefix is accepted. The paper's key evidence is an average accepted length of 5.0 out of 6 on a me

What carries the argument

Multi-token prediction (MTP): five Transformer branches that each predict one future transcript token (offset 1 to 5) in parallel with the main next-token head, sharing the embedding layer and vocabulary head. Each branch takes the previous branch's hidden state plus a shifted token embedding, then passes through a decoder-style block. The branches are trained first in isolation and then jointly with the decoder; at inference the model accepts the longest prefix of the six-token proposal that agrees with the main autoregressive path, so MTP acts as a safe acceleration primitive.

Load-bearing premise

The efficiency claim rests on computing expected accepted length as the sum of marginal acceptance rates (5.0), whereas the stated inference rule ('reject all later tokens once any prefix token fails') actually yields an expected accepted prefix of about 4.2; the ablation also assumes the pre-MTP model is the right control even though joint calibration updates the decoder.

What would settle it

Measure the actual number of tokens emitted per forward step on a meeting-speech benchmark under the deployed stop-at-first-rejection rule. If the empirical mean accepted prefix is substantially below 5.0, the paper's RTF and speed-up claims are overstated.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Per-step decoding cost drops to roughly a fifth of a full autoregressive step, giving a real-time factor of 0.0053 on a single GPU despite a 4B decoder.
  • The native 32K context window lets one pass cover up to 30 minutes of audio, avoiding the boundary errors of chunk-and-stitch pipelines.
  • Because the final transcript always follows the verified path, MTP preserves the accuracy of standard autoregressive decoding; the reported ablation shows average error-rate changes within 0.06 points.
  • Acceptance rates decay by roughly a factor of 0.9 per position, so the marginal benefit of adding branches shrinks; the paper picks MTP-5 as the efficiency–complexity sweet spot.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the expected accepted prefix is computed under the paper's own stop-at-first-rejection rule, the true average is about 4.2 tokens rather than 5.0, so the reported RTF advantage is likely optimistic; this is our calculation, not the paper's.
  • The MTP ablation's control (pre-MTP model) is not a pure counterfactual because joint calibration updates the whole decoder, so the claim that MTP alone is accuracy-neutral is not fully established by Table 4.
  • The anchored-generation argument should transfer to other input-constrained tasks such as speech translation or video captioning; a testable extension is to measure per-position acceptance rates there and compare them with ASR.
  • If speed is the priority, the verification threshold could be relaxed (e.g., accept non-contiguous matches with a fallback) or the branch horizon tuned per domain, potentially recovering some of the gap between 4.2 and 5.0.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. ParaASR proposes to accelerate LLM-based automatic speech recognition by adding Multi-Token Prediction (MTP) branches to a 4B-parameter decoder. At each decoding step the main branch predicts the next token and five auxiliary branches predict the following five tokens; a prefix-verification rule admits only the longest contiguous prefix that matches the autoregressive path. The authors report that this achieves an average accepted length of 5.0 out of 6 proposed tokens on a WenetSpeech meeting set, an RTF of 0.0053, and state-of-the-art or competitive WER/CER on Chinese, English, and long-form benchmarks. The training recipe is staged: audio-language pretraining, ASR supervised fine-tuning without MTP, then frozen-branch MTP alignment followed by joint calibration. The paper concludes that ASR is a particularly natural setting for multi-token decoding because speech anchors future tokens, and that decoder scaling, low latency, and long-context transcription need not be competing goals.

Significance. If the efficiency and safety claims were sustained, this would be a practically valuable result: it shows a concrete way to make LLM-based ASR faster without sacrificing accuracy, and the 32K-context single-pass long-form capability is useful. The paper also has comparative strengths: it evaluates on multiple external benchmarks, uses a relatively simple and reproducible architecture, and clearly separates the MTP training stages. However, the headline efficiency number appears to be computed with an incorrect formula under the paper's own decoding rule, and the MTP ablation is confounded by the joint-calibration stage. These issues directly affect the central claim that MTP enables 'safe acceleration' and that ASR is especially suited to multi-token prediction. The underlying idea remains plausible, but the evidence as presented does not currently support the quantitative conclusions.

major comments (2)
  1. [§2.2, Table 3] The reported 'Avg. Length' is computed as the sum of per-position acceptance rates, but under the paper's own prefix-verification rule the expected accepted prefix length is the sum of cumulative products. For MTP-5 the rates 0.95, 0.88, 0.80, 0.71, 0.64 give 1 + 0.95 + 0.836 + 0.669 + 0.475 + 0.304 ≈ 4.23, not 5.0; MTP-3 gives ≈3.48 not 3.6, and MTP-7 gives ≈4.57 not 6.1. The efficiency gain and the '39%/22%' diminishing-return comparison are therefore overstated. If the implementation actually accepts non-contiguous later tokens after an earlier rejection, then §2.2's equivalence to autoregressive decoding no longer holds and the 'safe acceleration' guarantee is void. Please report the actual prefix expectation or justify a different decoding rule.
  2. [§3.3, Table 4] The matched ablation is not matched. The 'w/o MTP' row is the model after ASR SFT but before MTP training, whereas the 'MTP-5' row includes the joint-calibration stage, which unfreezes the adapter and decoder and adds 10K further steps of optimization. Any accuracy change (e.g., VoxPopuli +0.46) could be due to this additional training rather than to MTP itself. To support the claim that MTP is a safe acceleration primitive, include a control that trains the same SFT model for the same number of steps without MTP branches, or freeze the backbone during MTP alignment.
minor comments (5)
  1. [Table 3] The notation 'Avg. Length 3.6 / 4', '5.0 / 6', '6.1 / 8' is ambiguous. State explicitly whether these are sums of marginal acceptance rates or expected prefix lengths under the verification rule; the two are different and the text currently conflates them.
  2. [§4.2, Table 3] The term 'strict per-position acceptance rate' should be defined: is it the probability that the h-th proposed token matches the autoregressive token conditional on all previous tokens being accepted, or an unconditional marginal rate? The formula used to derive average length depends on this.
  3. [References] Baseline 'Doubao-ASR-2603' is cited as reference [5], which is 'Seed-ASR'. These appear to be different systems; please provide the correct citation for Doubao-ASR-2603.
  4. [References] Reference [34] (a fetal ECG paper) is cited for the principle of combining multiple estimators before ROVER; this citation seems unrelated and should be replaced or removed.
  5. [Figure 2] The symbols 'H-Norm' and 'E-Norm' in the MTP block diagram are not defined in the caption or in the text. Please define them or remove the notation.

Circularity Check

0 steps flagged

No load-bearing circularity: the central quality and efficiency claims are measured on external benchmarks and local serving, not forced by construction. The paper's Table 3 'Avg. Length' sum is inconsistent with its own prefix-rejection rule, but that is an arithmetic/consistency problem rather than a circular derivation.

full rationale

The main claims of ParaASR are empirical and externally grounded. Recognition quality is reported on public benchmarks (AISHELL, LibriSpeech, Common Voice, FLEURS, VoxPopuli, Earnings22, WenetSpeech), and the RTF is measured under a local single-GPU serving setup, so the accuracy and latency results are not implied by the training objective or by any fitted parameter. The MTP design is an application of existing multi-token/speculative-decoding ideas to ASR, not a renamed version of the result it claims to establish. Self-citations to the authors' StepAudio series [21,30,31,41] provide the pretrained foundation and training recipe; these are antecedents of the current system, and the paper does not use them to forbid alternatives or to derive the MTP acceleration. The final model is independently evaluated relative to external baselines, so these self-citations are not load-bearing circularity. One non-circular concern deserves note: Table 3 reports 'Avg. Length 5.0/6' for MTP-5, which matches the simple sum 1 + 0.95 + 0.88 + 0.80 + 0.71 + 0.64, while Section 2.2 states that after a rejected proposal 'all subsequent proposed tokens are rejected.' Under that prefix rule the expected verified prefix length would be the cumulative-product sum, roughly 4.2, not 5.0. This is an internal-consistency or reporting issue in the efficiency evidence, not a circularity in the derivation chain, so it does not raise the circularity score above the minor-self-citation range.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central speed claim depends on H, α, and the pseudo-label quality threshold; the recognition/foundation quality depends on the frozen encoder, the unspecified pretrained text LLM, and the assumption that public test sets were not seen in SFT. None of these is independently evidenced in the paper.

free parameters (3)
  • MTP horizon H = 5
    Choice of five MTP branches selected after comparing MTP-3/5/7; reported as deliberate efficiency-complexity trade-off (§4.2, Table 3).
  • branch loss decay α = 0.9
    Exponential decay weight for MTP branch losses, set by hand in §3.3; not ablated.
  • long-form label disagreement threshold ê = 0.05
    Clips with normalized ROVER disagreement >5% are discarded in long-form pseudo-label pipeline (§3.2); chosen as quality filter, affects training supervision.
axioms (5)
  • domain assumption Audio-encoder features frozen during all stages remain sufficient for accurate ASR after adapter/decoder training.
    §2.1/§3.2 freeze the 0.6B encoder throughout; no unfreezing or encoder adaptation is tested.
  • domain assumption Five auxiliary heads sharing the backbone embedding/LM head can learn a valid future-token proposal distribution.
    §2.2/§3.3; the acceptance data is the only evidence, and its computation is in question.
  • domain assumption ROVER fusion of three ASR systems plus LLM refinement yields labels accurate enough for long-form training.
    §3.2 long-form pipeline; label quality is asserted, not verified against human transcripts.
  • standard math Speculative verification with prefix rejection preserves exact autoregressive outputs when draft and target distributions match.
    §2.2; standard speculative-decoding theorem, but the paper does not demonstrate the distribution match for its trained branches and its ablation shows nonzero transcript changes.
  • domain assumption Evaluation splits are disjoint from SFT training corpora.
    §3.2 says SFT integrates major public corpora; Table 1 evaluates on AISHELL/LibriSpeech/WenetSpeech, so exclusion of test partitions is assumed but unstated.

pith-pipeline@v1.3.0-daily-deepseek · 10341 in / 23919 out tokens · 251187 ms · 2026-08-03T10:07:57.504063+00:00 · methodology

0 comments
read the original abstract

Audio-encoder-LLM-decoder architectures have become the dominant paradigm for modern automatic speech recognition (ASR), improving transcription quality through large-scale language modeling. However, the cost of autoregressive decoding scales with decoder size, creating a fundamental trade-off between recognition quality and serving latency. We argue this trade-off is not inherent: unlike open-ended text generation, ASR outputs are strongly anchored to the input speech signal, providing a natural inductive bias toward high-parallelism decoding. Building on this, we introduce ParaASR, an ASR system that leverages Multi-Token Prediction (MTP) to let a 4B LLM decoder emit multiple tokens per forward step. Starting from a publicly available audio-language foundation, the model first establishes a robust autoregressive recognizer and then aligns five future-token branches through a staged optimization recipe. At inference, it proposes a six-token continuation per step and admits only the verified prefix into the transcript, preserving the safety of standard autoregressive decoding. The average accepted length reaches 5.0 out of 6 proposed tokens, confirming that the deterministic structure of speech makes ASR an especially natural setting for multi-token decoding. ParaASR further retains a native 32K-context window and transcribes up to 30 minutes of audio in a single pass. Across diverse benchmarks, it attains average error rates of 2.97%, 3.68%, and 3.70% on Chinese, English, and long-form evaluations, respectively, while reaching a real-time factor (RTF) as low as 0.0053. These results show that decoder scaling, low-latency inference, and long-context transcription need not be competing goals when future-token proposals are anchored by the acoustic signal and guarded by autoregressive verification.

Figures

Figures reproduced from arXiv: 2607.29279 by Bo Zhao, Daijiao Liu, Daxin Jiang, Fei Tian, Feng Tian, Gang Yu, Haoyang Zhang, Jinglan Gong, Jun Chen, Qingjian Lin, Xiangyu Tony Zhang, Xiangyu Zhang, Xie Li, Xuerui Yang, Yechang Huang, Yuxin Li, Yuxin Zhang.

Figure 1
Figure 1. Figure 1: Performance comparison across Chinese, English, and long-form ASR bench [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of ParaASR. A frozen audio encoder produces 80 ms acoustic em￾beddings, a linear adapter maps them to the decoder hidden space, and multiple MTP blocks propose future transcript tokens in parallel with the main next-token prediction. in ASR makes lookahead proposals remarkably predictable, effectively converting the deterministic structure of speech into a high-parallelism decoding advantage. … view at source ↗
Figure 3
Figure 3. Figure 3: Long-form ASR data construction pipeline. The process moves from individual [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 19 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2509.12508 (2025)

    An, K., Chen, Y., Chen, Z., Deng, C., Du, Z., Gao, C., et al.: Fun-ASR technical report. arXiv preprint arXiv:2509.12508 (2025)

    arXiv 2025

  2. [2]

    In: Proceedings of the Twelfth Language Resources and Evaluation Conference

    Ardila, R., Branson, M., Davis, K., et al.: Common voice: A massively-multilingual speech corpus. In: Proceedings of the Twelfth Language Resources and Evaluation Conference. pp. 4218–4222 (2020)

    2020

  3. [3]

    Artificial Analysis: Earnings22-cleaned-aa: Cleaned ground truth transcripts for earnings22 english test set (2026), https://artificialanalysis.ai/articles/aa-wer-v2

    2026

  4. [4]

    Artificial Analysis: Voxpopuli-cleaned-aa: Cleaned ground truth transcripts for voxpopuli english test set (2026), https://artificialanalysis.ai/articles/aa-wer-v2

    2026

  5. [5]

    arXiv preprint arXiv:2407.04675 (2024)

    Bai, Y., Chen, J., Chen, J., Chen, W., Chen, Z., Ding, C., Dong, L., Dong, Q., Du, Y., Gao, K., et al.: Seed-asr: Understanding diverse speech and contexts with llm-based speech recognition. arXiv preprint arXiv:2407.04675 (2024)

    Pith/arXiv arXiv 2024

  6. [6]

    arXiv preprint arXiv:2303.00747 (2023)

    Bain, M., Huh, J., Han, T., Zisserman, A.: Whisperx: Time-accurate speech tran- scription of long-form audio. arXiv preprint arXiv:2303.00747 (2023)

    Pith/arXiv arXiv 2023

  7. [7]

    In: 2017 20th Conference of the Oriental Chapter of the International Coordinating Committee on Speech Databases and Speech I/O Systems and Assessment

    Bu, H., Du, J., Na, X., Wu, B., Zheng, H.: AISHELL-1: An open-source mandarin speech corpus and a speech recognition baseline. In: 2017 20th Conference of the Oriental Chapter of the International Coordinating Committee on Speech Databases and Speech I/O Systems and Assessment. pp. 1–5 (2017)

    2017

  8. [8]

    arXiv preprint arXiv:2401.10774 (2024)

    Cai, T., Li, Y., Geng, Z., Peng, H., Lee, J.D., Chen, D., Dao, T.: Medusa: Simple llm inference acceleration framework with multiple decoding heads. arXiv preprint arXiv:2401.10774 (2024)

    Pith/arXiv arXiv 2024

  9. [9]

    arXiv preprint arXiv:1508.01211 (2015)

    Chan, W., Jaitly, N., Le, Q.V., Vinyals, O.: Listen, attend and spell. arXiv preprint arXiv:1508.01211 (2015)

    Pith/arXiv arXiv 2015

  10. [10]

    arXiv preprint arXiv:2302.01318 (2023)

    Chen, C., Borgeaud, S., Irving, G., Lespiau, J.B., Sifre, L., Jumper, J.: Acceler- ating large language model decoding with speculative sampling. arXiv preprint arXiv:2302.01318 (2023)

    Pith/arXiv arXiv 2023

  11. [11]

    arXiv preprint arXiv:2205.12446 (2022)

    Conneau, A., Ma, M., Khanuja, S., et al.: FLEURS: Few-shot learning evaluation of universal representations of speech. arXiv preprint arXiv:2205.12446 (2022)

    Pith/arXiv arXiv 2022

  12. [12]

    arXiv preprint arXiv:2511.00850 (2025)

    Deng, Y., Hu, G., Sun, H., Zhang, X., Zhang, H., Tian, F., Yang, X., Yu, G., Chng, E.S.: Multi-bench: A multi-turn interactive benchmark for assessing emotional intelligence ability of spoken dialogue models. arXiv preprint arXiv:2511.00850 (2025)

    arXiv 2025

  13. [13]

    arXiv preprint arXiv:1808.10583 (2018)

    Du, J., Na, X., Liu, X., Bu, H.: AISHELL-2: Transforming Mandarin ASR research into industrial scale. arXiv preprint arXiv:1808.10583 (2018)

    Pith/arXiv arXiv 2018

  14. [14]

    In: 1997 IEEE Workshop on Automatic Speech Recognition and Understanding Proceedings

    Fiscus, J.G.: A post-processing system to yield reduced word error rates: Recognizer output voting error reduction (ROVER). In: 1997 IEEE Workshop on Automatic Speech Recognition and Understanding Proceedings. pp. 347–354 (1997)

    1997

  15. [15]

    arXiv preprint arXiv:2404.19737 (2024)

    Gloeckle, F., Idrissi, B.Y., Roziere, B., Lopez-Paz, D., Synnaeve, G.: Better & faster large language models via multi-token prediction. arXiv preprint arXiv:2404.19737 (2024)

    Pith/arXiv arXiv 2024

  16. [16]

    In: Supervised sequence labelling with recurrent neural networks, pp

    Graves, A.: Connectionist temporal classification. In: Supervised sequence labelling with recurrent neural networks, pp. 61–93. Springer (2012)

    2012

  17. [17]

    arXiv preprint arXiv:1211.3711 (2012)

    Graves, A.: Sequence transduction with recurrent neural networks. arXiv preprint arXiv:1211.3711 (2012)

    Pith/arXiv arXiv 2012

  18. [18]

    arXiv preprint arXiv:2602.10604 (2026)

    Huang, A., Li, A., Kong, A., Wang, B., Jiao, B., Dong, B., Wang, B., Chen, B., Li, B., Ma, B., et al.: Step 3.5 Flash: Open frontier-level intelligence with 11B active parameters. arXiv preprint arXiv:2602.10604 (2026)

    arXiv 2026

  19. [19]

    In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Koluguri, N.R., Kriman, S., Zelenfroind, G., Majumdar, S., Rekesh, D., Noroozi, V., Balam, J., Ginsburg, B.: Investigating end-to-end asr architectures for long form audio transcription. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 13366–13370. IEEE (2024)

    2024

  20. [20]

    PMLR (2023)

    Leviathan, Y., Kalman, M., Matias, Y.: Fast inference from transformers via specu- lativedecoding.In:InternationalConferenceonMachineLearning.pp.19274–19286. PMLR (2023)

    2023

  21. [21]

    arXiv preprint arXiv:2605.23463 (2026)

    Lin, B., Zhao, B., Wu, B., Yan, C., Wu, C., Yi, C., Yao, C., Liu, D., Tian, F., Tian, F., et al.: StepAudio 2.5 technical report. arXiv preprint arXiv:2605.23463 (2026)

    Pith/arXiv arXiv 2026

  22. [22]

    arXiv preprint arXiv:2605.12034 (2026)

    Liu, C., Ma, L., Zhang, X.T., Zhang, Y., Zhang, H., Yang, X., Tian, F.: Boosting omni-modal language models: Staged post-training with visually debiased evalua- tion. arXiv preprint arXiv:2605.12034 (2026)

    Pith/arXiv arXiv 2026

  23. [23]

    arXiv preprint arXiv:2509.24310 (2025)

    Liu, H., Zhang, H., Zhang, Q., Zhang, X., Shi, D., Chng, E.S., Li, H.: Code- switching speech recognition under the lens: Model-and data-centric perspectives. arXiv preprint arXiv:2509.24310 (2025)

    arXiv 2025

  24. [24]

    Multimedia Tools and Applications80(6), 9411–9457 (2021)

    Malik, M., Malik, M.K., Mehmood, K., Makhdoom, I.: Automatic speech recogni- tion: a survey. Multimedia Tools and Applications80(6), 9411–9457 (2021)

    2021

  25. [25]

    In: 2015 IEEE International Conference on Acoustics, Speech and Signal Processing

    Panayotov, V., Chen, G., Povey, D., Khudanpur, S.: LibriSpeech: An ASR corpus based on public domain audio books. In: 2015 IEEE International Conference on Acoustics, Speech and Signal Processing. pp. 5206–5210 (2015)

    2015

  26. [26]

    In: Interspeech 2019

    Park, D.S., Chan, W., Zhang, Y., et al.: SpecAugment: A simple data augmentation method for automatic speech recognition. In: Interspeech 2019. pp. 2613–2617 (2019)

    2019

  27. [27]

    arXiv preprint arXiv:2601.18184 (2026)

    Peng, Z., Yu, J., Chang, Y., Wang, Z., Dong, L., Hao, Y., et al.: VIBEVOICE-ASR technical report. arXiv preprint arXiv:2601.18184 (2026)

    arXiv 2026

  28. [28]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International conference on machine learning. pp. 28492–28518. PMLR (2023)

    2023

  29. [29]

    arXiv preprint arXiv:2601.21337 (2026)

    Shi, X., Wang, X., Guo, Z., Wang, Y., Zhang, P., Zhang, X., et al.: Qwen3-ASR technical report. arXiv preprint arXiv:2601.21337 (2026)

    Pith/arXiv arXiv 2026

  30. [30]

    arXiv preprint arXiv:2511.15848 (2025)

    Tian,F.,Zhang,X.T.,Zhang,Y.,Zhang,H.,Li,Y.,Liu,D.,Deng,Y.,Wu,D.,Chen, J., Zhao, L., et al.: Step-Audio-R1 technical report. arXiv preprint arXiv:2511.15848 (2025)

    arXiv 2025

  31. [31]

    arXiv preprint arXiv:2507.16632 (2025)

    Wu, B., Yan, C., Hu, C., Yi, C., Feng, C., Tian, F., Shen, F., Yu, G., Zhang, H., Li, J., et al.: StepAudio 2 technical report. arXiv preprint arXiv:2507.16632 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    arXiv preprint arXiv:2510.09592 (2025)

    Wu, D., Zhang, H., Chen, J., Liu, H., Chng, E.S., Tian, F., Yang, X., Zhang, X., Jiang, D., Yu, G., et al.: Mind-paced speaking: A dual-brain approach to real-time reasoning in spoken language models. arXiv preprint arXiv:2510.09592 (2025)

    Pith/arXiv arXiv 2025

  33. [33]

    arXiv preprint arXiv:2509.17765 (2025)

    Xu, J., Guo, Z., Hu, H., Chu, Y., Wang, X., He, J., Wang, Y., Shi, X., He, T., Zhu, X., et al.: Qwen3-Omni technical report. arXiv preprint arXiv:2509.17765 (2025)

    Pith/arXiv arXiv 2025

  34. [34]

    In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Xuan, Y., Zhang, X., Li, S.S., Shen, Z., Xie, X., Garcia, L.P., Togneri, R.: A new approach to extract fetal electrocardiogram using affine combination of adaptive filters. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2023)

    2023

  35. [35]

    In: Proc

    Zeng, C., Wang, X., Miao, X., Cooper, E., Yamagishi, J.: Improving generaliza- tion ability of countermeasures for new mismatch scenario by combining multiple advanced regularization terms. In: Proc. Interspeech 2023. pp. 1998–2002 (2023)

    2023

  36. [36]

    In: ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing

    Zhang, B., Lv, H., Guo, P., Shao, Q., Yang, C., Xie, L., Xu, X., Bu, H., Chen, X., Zeng, C., et al.: WenetSpeech: A 10000+ hours multi-domain mandarin corpus for speech recognition. In: ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing. pp. 6182–6186 (2022)

    2022

  37. [37]

    arXiv preprint arXiv:2605.20755 (2026)

    Zhang, H., Chen, J., Wu, D., Li, Y., Zhang, Y., Zhang, X.T., Liu, C., Lin, Q., Peng, Y., Liu, H., et al.: DuplexSLA: A full-duplex spoken language model with synchronized speech, language, and action. arXiv preprint arXiv:2605.20755 (2026)

    Pith/arXiv arXiv 2026

  38. [38]

    arXiv preprint arXiv:2505.17076 (2025)

    Zhang, H., Liu, H., Zhang, X., Zhang, Q., Hu, Y., Zhao, J., Tian, F., Yang, X., Garcia, L.P., Chng, E.S.: Impact of frame rates on speech tokenizer: A case study on mandarin and english. arXiv preprint arXiv:2505.17076 (2025)

    Pith/arXiv arXiv 2025

  39. [39]

    arXiv preprint arXiv:2605.29209 (2026)

    Zhang, X., Li, Y., Zhang, H., Han, S., Liu, H., Zhang, Q., Ahmed, B., Epps, J.: The WER trap: Shattering the illusion of unified tokens in speech language models. arXiv preprint arXiv:2605.29209 (2026)

    Pith/arXiv arXiv 2026

  40. [40]

    IEEE Transactions on Audio, Speech and Language Processing (2025)

    Zhang, X., Zhang, Q., Liu, H., Xiao, T., Qian, X., Ahmed, B., Ambikairajah, E., Li, H., Epps, J.: Mamba in speech: Towards an alternative to self-attention. IEEE Transactions on Audio, Speech and Language Processing (2025)

    2025

  41. [41]

    arXiv preprint arXiv:2604.25719 (2026)

    Zhang, Y., Zhang, X.T., Liu, D., Tian, F., Deng, Y., Chen, J., Lin, Q., Zhang, H., Li, Y., Gong, J., et al.: Step-Audio-R1.5 technical report. arXiv preprint arXiv:2604.25719 (2026)

    Pith/arXiv arXiv 2026