Pith. sign in

REVIEW 3 major objections 5 minor 38 references

Causal Speech Enhancement with Predicting Semantics based on Quantized Self-supervised Learning Features

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that causal speech enhancement improves when a self-supervised-feature encoder is trained to predict future quantized semantic tokens, reaching 2.88 PESQ on VoiceBank+DEMAND.

desk verdict Plausible incremental causal SE paper whose central semantic-prediction claim rests on a 0.05 PESQ gap without error bars and with N chosen on the test set. read the letter →

arxiv 2412.19248 v1 pith:6TVPXGHW submitted 2024-12-26 eess.AS cs.SD

Emiru Tsunoo , Yuki Saito , Wataru Nakata , Hiroshi Saruwatari This is my paper · ORCID
classification eess.AScs.SD
keywords causalspeechenhancementself-supervisedlearningWavLMvectorquantizationsemantictokenpredictionmulti-taskfeature-wiselinearmodulationVoiceBank-DEMAND
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 a speech enhancer that is only allowed to see past audio can still exploit information about what will be said next. The authors' answer is yes: they take features from a self-supervised speech model computed with causal masking, quantize them into discrete phoneme-like tokens, and train the enhancer to also predict the next few tokens. They report a PESQ of 2.88 on the VoiceBank+DEMAND test set, and removing the semantic-prediction branch lowers PESQ to 2.83. The paper also claims this is the first method to fold causal SSL features into an enhancement model and to train the token prediction in one stage with vector quantization.

What carries the argument

The argument is carried by a causal semantic-token prediction loop inserted into a mask-estimating enhancement model. First, causal SSL features are obtained by running WavLM on each truncated prefix and taking the last frame output, a causal approximation of a full-context pretrained model. Second, a vector quantizer maps the weighted-sum causal features to discrete tokens with a 1024-entry codebook, trained jointly through moving-average updates and a commitment loss. Third, an SSL feature encoder takes the current causal features plus the current latent codebook vector and simultaneously predicts the next five tokens by minimizing cross-entropy, so the encoder becomes aware of phoneme continuation. The enhanced mask is then estimated by causal Transformer blocks conditioned through FiLM on this semantic-aware encoding.

What would settle it

Train the proposed model with the next-token prediction targets shuffled randomly across frames; if PESQ does not fall from 2.88 toward the 2.83 no-semantics baseline, the semantic-prediction branch is not the source of the gain.

Watch

Extended reading notes

Core claim

The central claim is that causal speech enhancement is improved when the model learns to predict future semantic content, represented as quantized self-supervised features. The proposed model computes WavLM features frame-by-frame using only the current and past input, combines them with log1p spectrogram features through feature-wise linear modulation, and estimates a speech mask. A separate branch quantizes the causal SSL features into a 1024-entry vector-quantized codebook and predicts the next five token identities during training. The paper reports 2.88 PESQ with this multi-task objective versus 2.83 without it, and shows that the gain largely disappears in a non-causal system where future context is already visible.

Load-bearing premise

The approach assumes that truncating a full-context pretrained speech model (WavLM) to past frames still yields representations informative enough to improve causal enhancement; the paper leaves the loss from this approximation unmeasured.

Editorial extensions

If this is right

  • Causal streaming enhancers can reuse off-the-shelf pretrained SSL models through causal truncation, gaining a 0.05 PESQ improvement from semantic prediction without retraining the SSL model.
  • The best prediction horizon is short: predicting five future tokens gives 2.88 PESQ, while eight or ten tokens drops to 2.83 and 2.82 as token accuracy falls.
  • Semantic prediction only helps when the future is genuinely hidden: in the non-causal system, with or without semantic prediction both give 3.14 PESQ.
  • Using the latent codebook vector as extra input outperforms both raw causal SSL features alone (2.80 PESQ) and the codebook index alone (2.86 PESQ).
  • A causal Transformer mask estimator outperforms a uni-directional LSTM by 0.06 PESQ in the same causal setting.

Reading between the lines

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

  • If the causal truncation were replaced by an SSL model actually trained under strict causality, the semantic-prediction gain might be larger, since the encoder would not need to compensate for features that originally assumed full-context attention.
  • The same strategy of predicting future semantic tokens as an auxiliary task could transfer to other causal audio restoration problems, such as dereverberation or packet-loss concealment, where phoneme continuation is similarly predictable.
  • Because the reported gain is 0.05 PESQ, a listener study or intelligibility test could reveal whether the semantic-prediction benefit is perceptually audible rather than only visible in an objective metric.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a causal speech enhancement (SE) system that combines causally restricted WavLM features with spectrogram features via FiLM, and augments the SE model with a vector-quantized semantic token predictor that predicts future tokens as a multi-task auxiliary loss. On VoiceBank+DEMAND, the reported PESQ is 2.88 with semantic prediction versus 2.83 without it. The authors also ablate input features, the number of predicted future tokens N, causal versus non-causal systems, FiLM versus concatenation, and causal Transformer versus LSTM. The main claimed contribution is that future semantic token prediction plays an important role in causal SE.

Significance. If the central claim is robust, the paper makes a useful contribution: it is, to my knowledge, the first to insert causal SSL features into an SE mask estimator, it integrates VQ tokenization in a one-stage training procedure, and it offers a clean ablation framework for semantic-continuation-aware causal SE. The architecture is clearly described and the ablation study is broad. However, the headline evidence for the semantic-prediction benefit is a single-run 0.05 PESQ difference with no confidence intervals, no significance testing, and a hyperparameter N chosen on the test set. The idea is plausible and the internal comparisons are mostly consistent, but the main quantitative claim needs stronger support before the paper can be accepted as a journal contribution.

major comments (3)
  1. [Abstract and §V-B, Table II] The central claim that semantic prediction 'played an important role' rests on the single-run comparison of PESQ 2.88 (with semantic prediction) versus 2.83 (without). No confidence intervals, multiple-seed results, or significance tests are reported. On VoiceBank+DEMAND, a 0.05 PESQ difference is often within run-to-run variability, especially with a small auxiliary loss weight of λ_ce=0.01 in Eq. (14). The non-causal control N1 vs N2 in Table IV is consistent with the proposed mechanism, but it does not by itself establish that the causal 0.05 gain is real. Please provide paired significance tests or repeated runs with variance reporting for the key with/without semantic prediction comparison.
  2. [§V-C, Table III] The number of predicted tokens N is selected by evaluating PESQ on the VoiceBank+DEMAND test set: the text states that 'we confirmed N=5 performed the best in the VoiceBank-DEMAND test set.' This is test-set selection, which can inflate the reported best value 2.88. The value N should be chosen on a held-out validation set (or via cross-validation), and the test-set result should be reported only for that fixed N. This is load-bearing because the 0.05 gain attributed to semantic prediction is partly the result of tuning N on the evaluation set.
  3. [§III-A, Eq. (5)] The causal approximation of WavLM, obtained by taking only the last frame of s(X≤t), is acknowledged to be a temporary solution ('It is desirable to train an SSL model that is particularly designed for a causal system, which remains for our future work'). However, the paper provides no analysis of how much information is lost or how the frame-wise restriction changes the SSL feature distribution. This limitation affects the interpretation of the absolute PESQ numbers and the 'causal SSL features' contribution, even though it does not undermine the internal with/without semantic-prediction comparison. A quantitative comparison of causal versus non-causal SSL feature quality, or an oracle upper-bound experiment, would make the first contribution more convincing.
minor comments (5)
  1. [§IV-A, Eqs. (7)-(10)] The semantic tokens are generated by VQ on the model's own causal SSL features, not by an independent phonetic labeler. The paper calls these 'semantic' based on prior work, but no evaluation of codebook quality or phoneme correlation is provided; consider adding an extrinsic measure or softening the semantic terminology.
  2. [Eq. (11)] The notation p(e_{t+1:t+N}|Z≤t) is slightly ambiguous because e_{t+1:t+N} is not formally defined as a tuple; please define it as (e_{t+1},...,e_{t+N}).
  3. [§II, Eq. (4) and §V-D] The phrase 'bi-directional long-short time memory (LSTM) layers' should be 'bi-directional long short-term memory (LSTM) layers.'
  4. [Experimental setup, §V-A] The paper would benefit from reporting training details such as number of runs, random seeds, batch size, number of parameters, and training time; no code or checkpoints are provided for independent verification.
  5. [Throughout] There are several spacing and formatting artifacts in the text, such as 'V oiceBank' in the Abstract, Introduction, and Conclusion; these should be corrected in the camera-ready version.

Circularity Check

2 steps flagged · score 4.0 of 10

The reported 0.05 PESQ gain from semantic-prediction MTL is partly circular because N=5 was chosen by maximizing PESQ on the VoiceBank+DEMAND test set, and the predicted 'semantic tokens' are the model's own VQ codes rather than independent phonetic labels.

  1. fitted input called prediction [Section V-C (Table III) combined with Section V-A/V-B and the Abstract]
    "As we increased the number of prediction N , the accuracy decreased because it is generally difficult to predict far future information. This accuracy drop adversely affect SE performance as PESQ decreased in the case of N = 8 and N = 10 comparing to the case of N = 5. Thus, we confirmed N = 5 performed the best in the VoiceBank-DEMAND test set."

    The headline PESQ of 2.88 is the maximum of the PESQ values reported for N in {1, 3, 5, 8, 10}, all measured on the same VoiceBank+DEMAND test set used for the final evaluation. N is therefore a hyperparameter selected to optimize the reported test metric, and the comparison '2.88 with semantic prediction versus 2.83 without' uses the N that was chosen to maximize PESQ on that same test set. The 0.05 gain is thus partly a selection artifact rather than an out-of-sample prediction. This is not a full equation-level equivalence, but it fits the fitted-input-called-prediction pattern.

  2. self definitional [Section IV-A (Eq. 7) and Section IV-B (Eqs. 11-12)]
    "The causal SSL features with weighted-sum (c(X)) are encoded by a linear layer as E(c(X)). Then the encoded features are quantized with codebooks by selecting the closest codebook e to the encoded output. The codebook is considered to be a discretized semantic token. ... We minimize the cross-entropy loss of the predicted token and the next tokens: Ł_ce = ... q(e_{t+n}) log p(e_{t+n}|Z≤t), where q(e) is the reference distribution given by applying VQ to the future frames."

    The 'semantic tokens' used as prediction targets are defined as the VQ codebook indices of the model's own causal SSL features c(X), which are themselves updated by the SE and VQ losses. Thus the auxiliary semantic-prediction objective is the model predicting its own self-generated codes, not predicting an independent phonetic label. The interpretation that this makes the model 'aware of phoneme continuation' therefore relies on the unverified assumption that these jointly-trained VQ codes correspond to phonemes. This is self-definitional in the terminology of 'semantic prediction', although it does not by itself force the measured PESQ improvement.

full rationale

The core SE evaluation is grounded in the external VoiceBank+DEMAND benchmark, and the main ablation (C2 versus C1 in Table IV) is a real architectural change, so the paper is not wholly circular. However, the central claimed evidence for the value of semantic prediction is weakened by two self-referential elements. First, the number of predicted tokens N is chosen by evaluating PESQ on the test set, so the headline 2.88 is the best of several test-set evaluations, making the 0.05 gap over the no-semantic baseline partly a selection artifact. Second, the semantic tokens are not external phonetic labels but the model's own VQ codes computed from the noisy causal SSL features, so 'semantic prediction' is a self-supervised objective whose phonetic interpretation is assumed rather than demonstrated. No load-bearing self-citation or imported uniqueness theorem was found; the citations to prior work by the same group (incremental TTS [25], run-and-back stitch search [31]) are technique and premise citations with independent corroboration. For these reasons the circularity is moderate: score 4 rather than 6-10, since the 0.05 gain is an empirical measurement rather than an equation-level reduction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the availability of useful causal SSL features from a non-causal pretrained model, the assumption that VQ tokens act as phonetic semantics, and hyperparameter choices, especially N=5, that are tuned on the test set. No new physical entities are introduced; the VQ codebook is a learned discrete representation rather than a new postulated object.

free parameters (4)
  • N (number of future semantic tokens predicted) = 5 (selected from {1,3,5,8,10})
    Chosen because it gave the highest PESQ (2.88) on the VoiceBank+DEMAND test set (Table III). The claimed semantic-prediction benefit depends on this selection.
  • lambda_ce (cross-entropy loss weight) = 0.01
    Hand-set to balance losses in Eq. (14). No sensitivity analysis is provided, and the ablation with and without semantic prediction uses this fixed weight.
  • codebook size = 1024
    Chosen as the VQ codebook size for semantic tokens; the paper does not motivate or vary this choice.
  • xi (VQ commitment balance) = 0.1
    Hand-set parameter in Eq. (7); no sensitivity analysis is reported.
assumptions (4)
  • ad hoc to paper WavLM, a non-causal pretrained model, remains informative when restricted to past frames only, as in Eq. (5).
    Section III-A defines causal SSL features by taking the last frame from s(X≤t). This is an approximation to make an inherently non-causal model causal; no measurement of information loss is provided, and the authors note a causal SSL model is future work.
  • domain assumption VQ-quantized SSL features represent phonetic or semantic units suitable for future prediction.
    Section IV-A adopts VQ to produce semantic tokens and cites [20,22] that SSL quantization correlates with phoneme states, but no evaluation of the learned codebook's phonetic meaning is given.
  • domain assumption Training the encoder to predict future semantic tokens improves the SE mask in a causal setting.
    Section IV-B motivates MTL by analogy to incremental TTS [25] and streaming VC [26]. The only evidence is the final PESQ ablation; the mechanism is not directly demonstrated.
  • domain assumption Masking the log1p noisy spectrogram and resynthesizing with the noisy phase is a valid enough SE framework for the comparisons.
    Equations (3) and (13) follow baseline [19]. The paper acknowledges that ignoring phase estimation is why it trails LFSFNet (Section V-B), so the comparison is restricted to magnitude-mask methods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Speech Enhancement with Predicting Semantics based on Quantized Self-supervised Learning Features." pith.science (2026). https://pith.science/paper/6TVPXGHW

@misc{pith2026241219248,
  author       = {Pith},
  title        = {Pith review of: Causal Speech Enhancement with Predicting Semantics based on Quantized Self-supervised Learning Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6TVPXGHW}},
  note         = {Machine review of arXiv:2412.19248}
}
read the original abstract

Real-time speech enhancement (SE) is essential to online speech communication. Causal SE models use only the previous context while predicting future information, such as phoneme continuation, may help performing causal SE. The phonetic information is often represented by quantizing latent features of self-supervised learning (SSL) models. This work is the first to incorporate SSL features with causality into an SE model. The causal SSL features are encoded and combined with spectrogram features using feature-wise linear modulation to estimate a mask for enhancing the noisy input speech. Simultaneously, we quantize the causal SSL features using vector quantization to represent phonetic characteristics as semantic tokens. The model not only encodes SSL features but also predicts the future semantic tokens in multi-task learning (MTL). The experimental results using VoiceBank + DEMAND dataset show that our proposed method achieves 2.88 in PESQ, especially with semantic prediction MTL, in which we confirm that the semantic prediction played an important role in causal SE.

Figures

Figures reproduced from arXiv: 2412.19248 by the authors.

Figure 1
Figure 1. SE model overview. C. Feature-wise linear modulation (FiLM) FiLM [33] is known to be an effective method to combine two modalities. While the original method concatenates log 1p features with SSL features, we apply FiLM to combine two features. The concatenation operation ⊕ in (3) is replaced by the FiLM manipulation, which is represented by ⊗ as X′ ⊗ c(X) = γ(c(X)) ⊙ α(X′ ) + β(c(X)), (6) where α, β, and γ are line… view at source ↗
Figure 2
Figure 2. Semantic token prediction model. The model predict n [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 34 canonical work pages

  1. [1]

    Speech enhancement based on deep denoising autoencoder,

    X. Lu et al. , “Speech enhancement based on deep denoising autoencoder,” in Proc. of Interspeech , 2013, pp. 436–440

    work page 2013

  2. [2]

    An experimental study on speech enhancement based on deep neural networks,

    Y . Xu et al. , “An experimental study on speech enhancement based on deep neural networks,” IEEE Signal processing letters, vol. 21, no. 1, pp. 65–68, 2013

    work page 2013

  3. [3]

    Multiple-target deep learning for LSTM- RNN based speech enhancement,

    L. Sun et al. , “Multiple-target deep learning for LSTM- RNN based speech enhancement,” in Proc. of HSCMA , 2017, pp. 136–140

    work page 2017

  4. [5]

    Phase-aware speech enhancement with deep complex U-net,

    H.-S. Choi et al., “Phase-aware speech enhancement with deep complex U-net,” in Proc. on ICLR , 2018

    work page 2018

  5. [6]

    U-net: Convolu - tional networks for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-net: Convolu - tional networks for biomedical image segmentation,” in Proc. of MICCAI , 2015, pp. 234–241

    work page 2015

  6. [7]

    DCCRN: Deep complex convolution recurrent network for phase-aware speech enhancement,

    Y . Hu et al. , “DCCRN: Deep complex convolution recurrent network for phase-aware speech enhancement,” in Proc. of Interspeech, 2020, pp. 2472–2476

    work page 2020

  7. [8]

    The interspeech 2020 deep noise suppres- sion challenge: Datasets, subjective testing framework, a nd challenge results,

    C. K. Reddy et al., “The interspeech 2020 deep noise suppres- sion challenge: Datasets, subjective testing framework, a nd challenge results,” in Proc. of Interspeech , 2020, pp. 2492– 2496

    work page 2020

  8. [9]

    MP-SENet: A speech enhancement model with parallel denoising of magnitude and phase spectra,

    Y .-X. Lu, Y . Ai, and Z.-H. Ling, “MP-SENet: A speech enhancement model with parallel denoising of magnitude and phase spectra,” in Proc. of Interspeech , 2023, pp. 3834–3838

    work page 2023

Show all 38 references
  1. [10]

    Conv-TasNet: Surpassing idea l time–frequency magnitude masking for speech separation,

    Y . Luo and N. Mesgarani, “Conv-TasNet: Surpassing idea l time–frequency magnitude masking for speech separation,” Transactions on Audio, Speech, and Language Processing , vol. 27, no. 8, pp. 1256–1266, 2019

    2019

  2. [11]

    A time–frequency smoothing neural network for speech enhancement,

    W. Y uan, “A time–frequency smoothing neural network for speech enhancement,” Speech Communication , vol. 124, pp. 75–84, 2020

    2020

  3. [12]

    A perceptually-motivated approach for low-complexity, real-time enhancement of fullband speech ,

    J.-M. V alin et al. , “A perceptually-motivated approach for low-complexity, real-time enhancement of fullband speech ,” in Proc. of Interspeech , 2020, pp. 2482–2486

    2020

  4. [13]

    DCCRN+: Channel-wise subband dccrn with snr estimation for speech enhancement,

    S. Lv et al. , “DCCRN+: Channel-wise subband dccrn with snr estimation for speech enhancement,” arXiv preprint arXiv:2106.08672, 2021

    2021 arXiv

  5. [14]

    Lightweight full-band and sub-ba nd fusion network for real time speech enhancement,

    Z. Chen and P . Zhang, “Lightweight full-band and sub-ba nd fusion network for real time speech enhancement,” in Proc. of Interspeech , 2022, pp. 921–925

    2022

  6. [15]

    OSSEM: One-shot speaker adaptive speech enhancement using meta learning,

    C. Y u et al. , “OSSEM: One-shot speaker adaptive speech enhancement using meta learning,” in Proc. of Interspeech , 2022, pp. 981–985

    2022

  7. [16]

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

    S. Chen et 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

  8. [17]

    Self- supervised learning for speech enhancement,

    Y .-C. Wang, S. V enkataramani, and P . Smaragdis, “Self- supervised learning for speech enhancement,” arXiv preprint arXiv:2006.10388, 2020

    2006 arXiv

  9. [18]

    Self-supervised denoising autoencoder with linear regression decoder for speech enhancement,

    R. E. Zezario et al. , “Self-supervised denoising autoencoder with linear regression decoder for speech enhancement,” in Proc. of ICASSP , 2020, pp. 6669–6673

    2020

  10. [19]

    Boosting self-supervised embeddings for speech enhancement,

    K.-H. Hung et al. , “Boosting self-supervised embeddings for speech enhancement,” in Proc. of Interspeech, 2022, pp. 186– 190

    2022

  11. [20]

    On generative spoken language modeling from raw audio,

    K. Lakhotia et al. , “On generative spoken language modeling from raw audio,” Transactions of the Association for Compu- tational Linguistics , vol. 9, pp. 1336–1354, 2021

    2021

  12. [21]

    Analysing discrete self super vised speech representation for spoken language modeling,

    A. Sicherman and Y . Adi, “Analysing discrete self super vised speech representation for spoken language modeling,” in Proc. of ICASSP , 2023, pp. 1–5

    2023

  13. [22]

    SELM: Speech enhancement using discrete tokens and language models,

    Z. Wang et al. , “SELM: Speech enhancement using discrete tokens and language models,” in Proc. of ICASSP , 2024, pp. 11 561–11 565

    2024

  14. [23]

    Uniaudio: An audio foundation model toward universal audio generation,

    D. Yang et al., “Uniaudio: An audio foundation model toward universal audio generation,” arXiv preprint arXiv:2310.00704, 2023

    2023 arXiv

  15. [24]

    Speechx: Neural codec language model as a versatile speech transformer,

    X. Wang et al. , “Speechx: Neural codec language model as a versatile speech transformer,” Transactions on Audio, Speech, and Language Processing , 2024

    2024

  16. [25]

    Low-latenc y incremental text-to-speech synthesis with distilled cont ext prediction network,

    T. Saeki, S. Takamichi, and H. Saruwatari, “Low-latenc y incremental text-to-speech synthesis with distilled cont ext prediction network,” in Proc. of ASRU , 2021, pp. 749–756

    2021

  17. [26]

    DualVC 3: Leveraging language model gen- erated pseudo context for end-to-end low latency streaming voice conversion,

    Z. Ning et al. , “DualVC 3: Leveraging language model gen- erated pseudo context for end-to-end low latency streaming voice conversion,” in Proc. of Interspeech , 2024

    2024

  18. [27]

    The voice bank corp us: Design, collection and data analysis of a large regional acc ent speech database,

    C. V eaux, J. Yamagishi, and S. King, “The voice bank corp us: Design, collection and data analysis of a large regional acc ent speech database,” in Proc. of O-COCOSDA/CASLRE , 2013, pp. 1–4

    2013

  19. [28]

    The diverse enviro n- ments multi-channel acoustic noise database (demand): A database of multichannel environmental noise recordings,

    J. Thiemann, N. Ito, and E. Vincent, “The diverse enviro n- ments multi-channel acoustic noise database (demand): A database of multichannel environmental noise recordings, ” in Proc. of Meetings on Acoustics , vol. 19, 2013

    2013

  20. [29]

    Investigating rnn-based speech enhancement methods for noise-robust text-to-speech.,

    C. V alentini-Botinhao et al. , “Investigating rnn-based speech enhancement methods for noise-robust text-to-speech.,” i n SSW, 2016, pp. 146–152

    2016

  21. [30]

    Boosting objective scores of a speech enhancement model by metricgan post-processing,

    S.-W. Fu et al. , “Boosting objective scores of a speech enhancement model by metricgan post-processing,” in Proc. of APSIPA ASC , 2020, pp. 455–459

    2020

  22. [31]

    Run-and-back stitch search: Novel block synchronous decoding for streaming encoder-decoder ASR,

    E. Tsunoo et al. , “Run-and-back stitch search: Novel block synchronous decoding for streaming encoder-decoder ASR,” in Proc of ICASSP , 2022, pp. 8287–8291

    2022

  23. [32]

    Attention is all you need,

    A. V aswani et al. , “Attention is all you need,” Advances in Neural Information Processing Systems , 2017

    2017

  24. [33]

    FiLM: Visual reasoning with a general conditioning layer,

    E. Perez et al. , “FiLM: Visual reasoning with a general conditioning layer,” in Proc. of AAAI , vol. 32, 2018

    2018

  25. [34]

    RepCodec: A speech representation codec for speech tokenization,

    Z. Huang, C. Meng, and T. Ko, “RepCodec: A speech representation codec for speech tokenization,” arXiv preprint arXiv:2309.00169, 2023

    2023 arXiv

  26. [35]

    Generating diverse high-fidelity images with VQ-V AE-2,

    A. Razavi, A. V an den Oord, and O. Vinyals, “Generating diverse high-fidelity images with VQ-V AE-2,” Advances in neural information processing systems , vol. 32, 2019

    2019

  27. [36]

    Autoregressive speech synthesis without vector quantization,

    L. Meng et al. , “Autoregressive speech synthesis without vector quantization,” arXiv preprint arXiv:2407.08551 , 2024

    2024 arXiv

  28. [37]

    Evaluation of objective quality measures for speech enhancement,

    Y . Hu and P . C. Loizou, “Evaluation of objective quality measures for speech enhancement,” Transactions on Audio, Speech, and Language Processing, vol. 16, no. 1, pp. 229–238, 2007

    2007

  29. [38]

    Exploring the best loss function for DNN-based low-latency speech enhancement with temporal convolutional networks,

    Y . Koyama et al. , “Exploring the best loss function for DNN-based low-latency speech enhancement with temporal convolutional networks,” arXiv preprint arXiv:2005.11611 , 2020

    2005 arXiv

  30. [39]

    A convolutional recurrent neural network for real-time speech enhancement,

    K. Tan and D. Wang, “A convolutional recurrent neural network for real-time speech enhancement,” in Proc. of In- terspeech, 2018, pp. 3229–3233

    2018

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.