Pith. sign in

REVIEW 2 major objections 5 minor 34 references

CNVSRC 2024: The Second Chinese Continuous Visual Speech Recognition Challenge

T0 review · 2 major / 5 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 CNVSRC 2024 reports that a baseline using Chinese character units and a Bi-Transformer decoder cuts character error rate on identical evaluation sets from 48.60% to 39.66% for single-speaker and from 58.37% to 52.20% for multi-speaker…

desk verdict A solid challenge report with a stronger Chinese VSR baseline and a new ~200-hour dataset; the baseline gain is real but the paper over-attributes it to two model changes without ablating other training differences. read the letter →

arxiv 2506.02010 v1 pith:QFDZK2SB submitted 2025-05-27 cs.CV cs.SDeess.AS

Zehua Liu , Xiaolou Li , Chen Chen , Lantian Li , Dong Wang This is my paper
classification cs.CVcs.SDeess.AS
keywords visualspeechrecognitionlipreadingChineselarge-vocabularyVSRCNVSRCchallengecharactererrorrateBi-TransformerdecoderunitshybridCTC/Attention
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 presents the second Chinese Continuous Visual Speech Recognition Challenge, a shared benchmark for large-vocabulary lip reading from silent face videos. Its central contribution is a stronger public baseline: the organizers report that switching from subword units to Chinese character units and replacing the standard forward Transformer decoder with a Bi-Transformer decoder lowers the character error rate on the same evaluation sets from 48.60% to 39.66% for single-speaker visual speech recognition and from 58.37% to 52.20% for the multi-speaker task. The challenge also releases an additional open-track dataset, CN-CVS2-P1, and the best participant systems outperform the baseline by a wide margin. A sympathetic reader would take the paper to establish that these two changes are enough to deliver a major, reproducible gain in Chinese large-vocabulary visual speech recognition.

What carries the argument

The load-bearing mechanism is the pairing of character-level output units with a Bi-Transformer decoder inside a hybrid CTC/Attention architecture. The decoder places a 3-layer reverse Transformer beside a 6-layer forward Transformer on top of a Conformer (convolution-augmented Transformer) encoder; during training both streams produce cross-entropy losses, which lets the forward decoder learn from backward context, and during inference the reverse stream is removed and only the forward Transformer, weighted 0.7 against CTC's 0.3, performs one-pass beam search. The character units change the prediction vocabulary from subword pieces to individual Chinese characters, which the authors argue gives the model a more natural target for the compositional structure of Chinese text. These components are trained with a joint CTC/Attention loss and a two-stage pretraining-then-fine-tuning schedule on the CN-CVS corpus.

What would settle it

Run the baseline in four configurations that swap one factor at a time: subword units with a standard forward decoder, character units with a standard decoder, subword units with the Bi-Transformer decoder, and character units with the Bi-Transformer decoder, holding all other training details fixed. If the full configuration does not beat each single-change variant by a clear margin on the same development and evaluation sets, the paper's attribution would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a baseline built on the previous year's architecture can be made substantially better by two targeted changes: modeling output with 4,468 Chinese character units instead of 5,904 subword units, and implementing the Attention decoder as a Bi-Transformer with a 3-layer reverse Transformer running in parallel with a 6-layer forward Transformer. On the single-speaker evaluation set the character error rate falls from 48.60% to 39.66%; on the multi-speaker set it falls from 58.37% to 52.20%. Because the reverse stream is discarded at inference, the improvement comes with no extra decoding cost. The paper further reports that the winning submissions, using augmented preprocessing, audio-visual pretraining and fusion, state-space decoders, and system fusion, push the error rate down to roughly 30% on the same tracks.

Load-bearing premise

The paper's main claim depends on the assumption that the named modeling changes, not other differences in the training recipe, are what lowered the error rate; no experiment is reported that isolates those changes from the rest of the setup.

Editorial extensions

If this is right

  • Because CNVSRC 2024 reuses the exact evaluation protocol and metric of CNVSRC 2023, the reported CER drop is a direct comparison on identical test sets rather than an artifact of changed benchmarks.
  • Since the reverse Transformer is discarded at inference, other lip-reading systems can adopt the same training-side trick without adding runtime latency.
  • The new CN-CVS2-P1 open-track dataset adds roughly 197 hours of Internet-sourced video, so open-track systems can be trained with substantially more data than fixed-track systems.
  • The gap between the baseline (39.66% single-speaker CER) and the best submissions (30.06% single-speaker CER) shows that the challenge's improvements leave considerable room for participant techniques.
  • With the best CER still above 30%, the authors conclude that Chinese large-vocabulary visual speech recognition remains far from real-world use and point to linguistic constraints from large language models as the next likely source of progress.

Reading between the lines

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

  • An editorial caution: the paper credits the baseline gain to character units and the Bi-Transformer, but reports no ablation that isolates these two changes from other training-protocol differences—such as a re-randomized classification layer, different learning rates, and a different checkpoint-averaging window—so part or all of the gain could come from those confounds.
  • A clean 2x2 experiment that toggles each factor separately would settle which change actually carries the improvement; a reader should not assume the architecture is the cause until that is run.
  • The strong showing of S4D decoders and KL-consistency losses among participants suggests decoder-side context modeling, not just the visual frontend, is where Chinese VSR progress is currently concentrated.
  • Because open-track data now differs from fixed-track data, future leaderboard comparisons should control for data volume; otherwise modeling advances and extra training data are conflated.
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

2 major / 5 minor

Summary. This paper presents the second Chinese Continuous Visual Speech Recognition Challenge, describing two evaluation tasks (single-speaker and multi-speaker VSR), the fixed/open track setup, the data profile, and the organizers' baseline systems. The baseline is built on the Auto-AVSR framework and incorporates two stated changes relative to the CNVSRC 2023 baseline: Chinese character units instead of subword units, and a Bi-Transformer attention decoder. The reported CERs on the held-out evaluation sets decrease from 48.60% to 39.66% (T1) and 58.37% to 52.20% (T2) relative to the 2023 baseline. The paper also releases an additional dataset CN-CVS2-P1 for open tracks and summarizes the techniques of the 10 submitted systems from 4 teams, with best CERs of 30.06% (T1 open), 30.46% (T1 fixed), and 34.29% (T2 fixed/open).

Significance. The challenge delivers a valuable, reproducible benchmark: the evaluation labels stayed confidential, the baseline is described in enough detail to reimplement, and the datasets, models, and code are open-sourced. If the reported baseline improvements are reliable, they provide a stronger public reference point for Chinese LVC-VSR. However, the paper's central attribution of the improvement to character units and the Bi-Transformer decoder is not supported by ablations, as multiple other training-protocol changes were introduced simultaneously. This limits the scientific claim, though the leaderboard results themselves remain informative.

major comments (2)
  1. [Section 3 (opening), Table 4; confounds in Section 3.3/Table 3] The paper claims that the two stated changes—Chinese character units and a Bi-Transformer decoder—'have yielded a substantial performance boost' (Section 3) and Table 4 shows large reductions (T1 Eval 48.60→39.66; T2 Eval 58.37→52.20). However, Section 3.3 and Table 3 document at least four additional differences relative to the CNVSRC 2023 baseline that are not ablated: (i) re-randomization of the classification layer in single-speaker fine-tuning, (ii) different learning rates across tasks and stages (0.0003 vs 0.0002 for FT; 0.0002/0.001 for P1/P2), (iii) different checkpoint averaging schemes (top-10 vs last-10 vs last-5), and (iv) the 2023 baseline's exact pretraining/fine-tuning recipe is not fully specified. Without ablations that isolate character units and the Bi-Transformer decoder from these confounds, the causal attribution is unsupported. Please add ablation experiments or revise the claim to state that the combined modification set yields the improvement.
  2. [Section 3.3 (Fine-tuning)] The fine-tuning description is internally inconsistent: it first states 'This process is the same for the models trained for the single-speaker and multi-speaker tasks' and then immediately lists differences (learning rate and re-randomized classification layer). It is therefore unclear whether the classification-layer re-randomization applies only to T1 or also to T2. Since this affects the reproducibility of the T2 baseline, please rewrite the paragraph to give the exact procedure for both tasks.
minor comments (5)
  1. [Section 1] 'mainly primarily' contains a doubled adverb; please fix to one.
  2. [Table 5] The leaderboard is missing Rank3 entries for T1 Open and T2 Open; please indicate whether there were fewer than three submissions in those tracks.
  3. [Section 4.1] 'All these teams significantly outperformed the baseline systems' uses 'significantly' without statistical tests; given the large gaps, I suggest 'substantially' or adding confidence intervals.
  4. [Section 4.2.2] 'A V-HuBERT' contains an incorrect space; it should be 'AV-HuBERT'.
  5. [Section 3.4] Please provide the version or citation for the TorchMetrics tool used for CER computation.

Circularity Check

0 steps flagged · score 2.0 of 10

Evaluation report with held-out benchmark results; no derivation-level circularity, only a minor self-citation pattern.

full rationale

This paper is a challenge summary, not a derivation. The baseline CERs in Table 4 (T1 Eval 48.60%->39.66%, T2 Eval 58.37%->52.20%) are measured on confidential held-out evaluation sets, not fitted to them, and the leaderboard results come from independent participant submissions. The paper relies on the authors' own prior work for the dataset (CN-CVS, ref [14]) and the base framework (Auto-A VSR, ref [16]), and it cites the previous CNVSRC 2023 report [15] for the preprocessing pipeline and the 2023 baseline numbers. This is a self-citation pattern, but it is not load-bearing for the central measured comparison: the evaluation labels are external and the CER metric (Eq. 1) is a standard definition applied to system outputs, not an equation that reduces to its own inputs. The stated attribution that character units and the Bi-Transformer decoder 'have yielded a substantial performance boost' is an empirical claim that is weakened by unablated training-protocol differences (Section 3.3), but that is a causal-inference gap, not circularity. No step in the paper defines, fits, or derives its conclusions from the quantities it claims to predict.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The challenge's central comparisons rest on three assumptions: identical evaluation protocol to the previous challenge, confidentiality of the eval labels, and representativeness of the eval sets. No new theoretical entities are introduced; CN-CVS2-P1 is a data artifact, not a postulated entity. The hand-chosen hyperparameters listed above affect the baseline CER but are not fitted to the eval set.

free parameters (2)
  • CTC/Attention beam-search interpolation weight = 0.7 (decoder), 0.3 (CTC)
    Hand-chosen in Section 3.1; directly affects the reported CER and is not ablated.
  • Development set fine-tuning split ratio = 8:1 (T1), 3:1 (T2)
    Hand-chosen in Section 3.3; different splits change the amount of adaptation data, influencing baseline CER comparisons.
assumptions (3)
  • domain assumption The CNVSRC 2024 evaluation protocol, data splits, and CER metric are identical to CNVSRC 2023.
    Stated in Section 2.1; if the eval sets or metric changed, the reduction from 48.60% to 39.66% would not be a valid comparison.
  • domain assumption Participants could not access the confidential text and audio of the evaluation sets during the competition.
    Stated in Section 2.2; the integrity of the leaderboard CERs depends on this confidentiality being honored.
  • domain assumption The evaluation sets are representative of the two target scenarios: recording-studio reading and Internet speech.
    Section 2.2 describes the data sources; if the eval samples are not representative, conclusions about real-world progress are weakened.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CNVSRC 2024: The Second Chinese Continuous Visual Speech Recognition Challenge." pith.science (2026). https://pith.science/paper/QFDZK2SB

@misc{pith2026250602010,
  author       = {Pith},
  title        = {Pith review of: CNVSRC 2024: The Second Chinese Continuous Visual Speech Recognition Challenge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QFDZK2SB}},
  note         = {Machine review of arXiv:2506.02010}
}
read the original abstract

This paper presents the second Chinese Continuous Visual Speech Recognition Challenge (CNVSRC 2024), which builds on CNVSRC 2023 to advance research in Chinese Large Vocabulary Continuous Visual Speech Recognition (LVC-VSR). The challenge evaluates two test scenarios: reading in recording studios and Internet speech. CNVSRC 2024 uses the same datasets as its predecessor CNVSRC 2023, which involves CN-CVS for training and CNVSRC-Single/Multi for development and evaluation. However, CNVSRC 2024 introduced two key improvements: (1) a stronger baseline system, and (2) an additional dataset, CN-CVS2-P1, for open tracks to improve data volume and diversity. The new challenge has demonstrated several important innovations in data preprocessing, feature extraction, model design, and training strategies, further pushing the state-of-the-art in Chinese LVC-VSR. More details and resources are available at the official website.

Figures

Figures reproduced from arXiv: 2506.02010 by the authors.

Figure 1
Figure 1. The model structure of CNVSRC 2024 baselines. Note that the reverse Transformer is discarded during infer￾ence; hence, it is drawn in grey. 3.2. Data Preprocessing The videos of the provided datasets contain the entire faces of target speakers; thus, a preprocessing pipeline was employed to extract the lip region. This pipeline was applied to both the training data (CN-CVS) and the development/evaluation data (CNVSR… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 27 canonical work pages

  1. [1]

    It has a wide range of applications, including public surveillance, assistive technologies for the elderly and disabled, and deepfake detection [1, 2, 3, 4]

    Introduction Visual Speech Recognition (VSR), also known as lip reading, is a technology that deciphers speech content from lip move- ments. It has a wide range of applications, including public surveillance, assistive technologies for the elderly and disabled, and deepfake detection [1, 2, 3, 4]. Despite significant advance- ments in word- and phrase-lev...

    work page 2023

  2. [2]

    CNVSRC 2024: The Second Chinese Continuous Visual Speech Recognition Challenge

    Tasks and Data The tasks in CNVSRC 2024 are the same as those in CNVSRC 2023 [15], mainly divided into two categories: 1) tasks for a specific speaker scenario, and 2) tasks for non-specific speakers. Each scenario is further divided into two tracks: open track and fixed track. To further support model development for the open track, CNVSRC 2024 introduce...

    work page Pith review arXiv 2024

  3. [3]

    Only the datasets provided by the organizer (except CN-CVS2-P1) were used, meaning these systems conform to the specifications of the fixed tracks

    Baseline System We developed two baseline systems based on the Auto-A VSR framework [16]: one for the Single-speaker VSR task (T1) and the other for the Multi-speaker VSR task (T2). Only the datasets provided by the organizer (except CN-CVS2-P1) were used, meaning these systems conform to the specifications of the fixed tracks. Building on the technical i...

    work page 2023

  4. [4]

    Leaderboard CNVSRC 2024 received 10 valid submissions from 4 partic- ipating teams

    CNVSRC 2024 Report 4.1. Leaderboard CNVSRC 2024 received 10 valid submissions from 4 partic- ipating teams. Notably, in contrast to CNVSRC 2023, more teams this year focused on the multi-speaker track, indicating that VSR research is gradually shifting from single-speaker sce- narios to more challenging multi-speaker scenarios. This shift highlights the g...

    work page 2024

  5. [5]

    Compared to CNVSRC 2023, this year’s challenge introduced a stronger baseline system and more train- ing data, further pushing the boundary of the Chinese VSR re- search

    Conclusion This paper presents a comprehensive overview of the sec- ond Chinese Continuous Visual Speech Recognition Challenge (CNVSRC 2024), highlighting its contributions to advancing Chinese LVC-VSR. Compared to CNVSRC 2023, this year’s challenge introduced a stronger baseline system and more train- ing data, further pushing the boundary of the Chinese...

    work page 2024

  6. [6]

    Continuous automatic speech recognition by lipreading,

    A. J. Goldschen, O. N. Garcia, and E. D. Petajan, “Continuous automatic speech recognition by lipreading,” in Motion-Based recognition. Springer, 1997, pp. 321–343

    work page 1997

  7. [7]

    Audiovisual in- tegration and lipreading abilities of older adults with normal and impaired hearing,

    N. Tye-Murray, M. S. Sommers, and B. Spehar, “Audiovisual in- tegration and lipreading abilities of older adults with normal and impaired hearing,” Ear and hearing, vol. 28, no. 5, pp. 656–668, 2007

    work page 2007

  8. [8]

    Visual speech recognition,

    A. B. Hassanat, “Visual speech recognition,” Speech and Lan- guage Technologies, vol. 1, pp. 279–303, 2011

    work page 2011

Show all 34 references
  1. [9]

    Lips don’t lie: A generalisable and robust approach to face forgery de- tection,

    A. Haliassos, K. V ougioukas, S. Petridis, and M. Pantic, “Lips don’t lie: A generalisable and robust approach to face forgery de- tection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 5039–5049

    2021

  2. [10]

    Multi-temporal lip-audio mem- ory for visual speech recognition,

    J. H. Yeo, M. Kim, and Y . M. Ro, “Multi-temporal lip-audio mem- ory for visual speech recognition,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  3. [11]

    Training strategies for improved lip-reading,

    P. Ma, Y . Wang, S. Petridis, J. Shen, and M. Pantic, “Training strategies for improved lip-reading,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2022, pp. 8472–8476

    2022

  4. [12]

    Improved word-level lipreading with temporal shrinkage net- work and netvlad,

    H. Yang, T. Luo, Y . Zhang, M. Song, L. Xie, Y . Yan, and E. Yin, “Improved word-level lipreading with temporal shrinkage net- work and netvlad,” inProceedings of the 2022 International Con- ference on Multimodal Interaction, 2022, pp. 504–508

    2022

  5. [13]

    Lip reading in the wild,

    J. S. Chung and A. Zisserman, “Lip reading in the wild,” in Com- puter Vision–ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. Springer, 2017, pp. 87–103

    2016

  6. [14]

    Distinguishing homophenes using multi-head visual-audio memory for lip reading,

    M. Kim, J. H. Yeo, and Y . M. Ro, “Distinguishing homophenes using multi-head visual-audio memory for lip reading,” in Pro- ceedings of the AAAI conference on artificial intelligence, vol. 36, no. 1, 2022, pp. 1174–1182

    2022

  7. [15]

    Con- former is all you need for visual speech recognition,

    O. Chang, H. Liao, D. Serdyuk, A. Shahy, and O. Siohan, “Con- former is all you need for visual speech recognition,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 10 136– 10 140

    2024

  8. [16]

    Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video,

    D. Serdyuk, O. Braga, and O. Siohan, “Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video,” in INTERSPEECH, 2022, pp. 2833–2837

    2022

  9. [17]

    Learning audio-visual speech representation by masked multimodal cluster prediction,

    B. Shi, W.-N. Hsu, K. Lakhotia, and A. Mohamed, “Learning audio-visual speech representation by masked multimodal cluster prediction,” in International Conference on Learning Representa- tions

  10. [18]

    Large- scale visual speech recognition,

    B. Shillingford, Y . Assael, M. W. Hoffman, T. Paine, C. Hughes, U. Prabhu, H. Liao, H. Sak, K. Rao, L. Bennett et al., “Large- scale visual speech recognition,” in INTERSPEECH, 2019, pp. 4135–4139

    2019

  11. [19]

    CN-CVS: A Mandarin audio-visual dataset for large vocabulary continuous visual to speech synthesis,

    C. Chen, D. Wang, and T. F. Zheng, “CN-CVS: A Mandarin audio-visual dataset for large vocabulary continuous visual to speech synthesis,” inICASSP 2023-2023 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

    2023

  12. [20]

    CNVSRC 2023: The first Chinese continuous visual speech recognition challenge,

    C. Chen, Z. Liu, X. Li, L. Li, and D. Wang, “CNVSRC 2023: The first Chinese continuous visual speech recognition challenge,” in INTERSPEECH, 2024, pp. 1930–1934

    2023

  13. [21]

    Auto-A VSR: Audio-visual speech recognition with automatic labels,

    P. Ma, A. Haliassos, A. Fernandez-Lopez, H. Chen, S. Petridis, and M. Pantic, “Auto-A VSR: Audio-visual speech recognition with automatic labels,” in ICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  14. [22]

    Combining residual net- works with LSTMs for lipreading,

    T. Stafylakis and G. Tzimiropoulos, “Combining residual net- works with LSTMs for lipreading,” in INTERSPEECH, 2017, pp. 3652–3656

    2017

  15. [23]

    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 et al. , “Conformer: Convolution- augmented Transformer for speech recognition,” in INTER- SPEECH, 2020, pp. 5036–5040

    2020

  16. [24]

    U2++: Unified two-pass bidirectional end-to-end model for speech recognition,

    D. Wu, B. Zhang, C. Yang, Z. Peng, W. Xia, X. Chen, and X. Lei, “U2++: Unified two-pass bidirectional end-to-end model for speech recognition,” arXiv preprint arXiv:2106.05642, 2021

    2021 arXiv

  17. [25]

    Hybrid CTC/attention architecture for end-to-end speech recog- nition,

    S. Watanabe, T. Hori, S. Kim, J. R. Hershey, and T. Hayashi, “Hybrid CTC/attention architecture for end-to-end speech recog- nition,” IEEE Journal of Selected Topics in Signal Processing , vol. 11, no. 8, pp. 1240–1253, 2017

    2017

  18. [26]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” in International Conference on Learning Representations , 2019

    2019

  19. [27]

    The NPU- ASLP-LiAuto system description for visual speech recognition in CNVSRC 2023,

    H. Wang, P. Guo, W. Chen, P. Zhou, and L. Xie, “The NPU- ASLP-LiAuto system description for visual speech recognition in CNVSRC 2023,” arXiv preprint arXiv:2401.06788, 2024

    2023 arXiv

  20. [28]

    Enhancing lip reading with multi-scale video and multi-encoder,

    H. Wang, P. Guo, X. Wan, H. Zhou, and L. Xie, “Enhancing lip reading with multi-scale video and multi-encoder,”arXiv preprint arXiv:2404.05466, 2024

    2024 arXiv

  21. [29]

    MFA-Conformer: Multi-scale feature aggregation con- former for automatic speaker verification,

    Y . Zhang, Z. Lv, H. Wu, S. Zhang, P. Hu, Z. Wu, H.-y. Lee, and H. Meng, “MFA-Conformer: Multi-scale feature aggregation con- former for automatic speaker verification,” in INTERSPEECH, 2022, pp. 306–310

    2022

  22. [30]

    E-branchformer: Branchformer with enhanced merging for speech recognition,

    K. Kim, F. Wu, Y . Peng, J. Pan, P. Sridhar, K. J. Han, and S. Watanabe, “E-branchformer: Branchformer with enhanced merging for speech recognition,” in 2022 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2023, pp. 84–91

    2022

  23. [31]

    On the parameteriza- tion and initialization of diagonal state space models,

    A. Gu, K. Goel, A. Gupta, and C. R ´e, “On the parameteriza- tion and initialization of diagonal state space models,” Advances in Neural Information Processing Systems , vol. 35, pp. 35 971– 35 983, 2022

    2022

  24. [32]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,” in International Conference on Learning Representations, 2022

    2022

  25. [33]

    Structured state space decoder for speech recognition and synthesis,

    K. Miyazaki, M. Murata, and T. Koriyama, “Structured state space decoder for speech recognition and synthesis,” in ICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  26. [34]

    A post-processing system to yield reduced word er- ror rates: Recognizer output voting error reduction (rover),

    J. G. Fiscus, “A post-processing system to yield reduced word er- ror rates: Recognizer output voting error reduction (rover),” in 1997 IEEE Workshop on Automatic Speech Recognition and Un- derstanding Proceedings. IEEE, 1997, pp. 347–354

    1997

Pith tools

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