pith. machine review for the scientific record. sign in

arxiv: 2510.06201 · v2 · submitted 2025-10-07 · 📡 eess.AS · cs.AI· cs.CL· cs.LG· cs.SD

TokenChain: A Discrete Speech Chain via Semantic Token Modeling

Mingxuan Wang , Satoshi Nakamura This is my paper

Pith reviewed 2026-05-18 09:12 UTC · model grok-4.3

classification 📡 eess.AS cs.AIcs.CLcs.LGcs.SD
keywords semantic tokensspeech chainASRTTSdiscrete modelingjoint trainingGumbel-Softmaxstraight-through estimator
0
0 comments X

The pith

TokenChain closes the speech chain loop using discrete semantic tokens to let ASR and TTS train each other.

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

The paper introduces TokenChain, a fully discrete machine speech chain that couples semantic-token ASR with a two-stage TTS model. An autoregressive text-to-semantic component is co-trained with ASR, while a masked-generative semantic-to-acoustic model handles final synthesis. Straight-through argmax and Gumbel-Softmax estimators pass gradients across the discrete token interface, balanced against supervised ASR loss via dynamic weight averaging. On LibriSpeech this yields faster convergence and lower equal-epoch error; on TED-LIUM it cuts relative ASR WER by 56 percent and T2S WER by 31 percent with minimal forgetting. The results indicate that chain learning remains effective when the perception-production loop operates entirely on semantic tokens rather than continuous waveforms.

Core claim

TokenChain models the human perception-production loop as a closed discrete chain: semantic-token ASR predicts tokens from audio while a co-trained autoregressive text-to-semantic model and a separate masked semantic-to-acoustic synthesizer complete the loop. End-to-end feedback flows through straight-through argmax and Gumbel-Softmax estimators, enabling the joint system to surpass baseline accuracy 2-6 epochs earlier, achieve 5-13 percent lower equal-epoch error on LibriSpeech, and reduce relative WER by 56 percent for ASR and 31 percent for T2S on TED-LIUM with little forgetting.

What carries the argument

The discrete semantic token interface that links ASR to the autoregressive text-to-semantic TTS stage, with straight-through argmax and Gumbel-Softmax estimators transmitting gradients across the non-differentiable token boundary.

If this is right

  • Chain learning stays effective when perception and production are linked by semantic tokens rather than continuous signals.
  • Dynamic weight averaging between chain feedback and supervised ASR loss stabilizes joint training.
  • Temperature schedules for Gumbel-Softmax control the balance between in-domain and cross-domain transfer.
  • The two-stage TTS design keeps text-to-speech synthesis stable while ASR improves.

Where Pith is reading between the lines

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

  • The same token interface could be tested on paired tasks such as speech translation or voice conversion to see whether chain benefits generalize.
  • Replacing continuous audio with tokens may lower memory and compute demands when scaling larger joint models.
  • Adding a third token stage for prosody or emotion could extend the chain without breaking the discrete gradient path.

Load-bearing premise

Straight-through argmax and Gumbel-Softmax estimators can transmit useful gradient signals across the discrete semantic-token interface without causing excessive instability or information loss during joint ASR-TTS training.

What would settle it

If identical models trained without the discrete token interface or with continuous acoustic features instead show no gain in convergence speed or error reduction, the benefit attributed to the token-based chain would be falsified.

read the original abstract

Machine Speech Chain, simulating the human perception-production loop, proves effective in jointly improving ASR and TTS. We propose TokenChain, a fully discrete speech chain coupling semantic-token ASR with a two-stage TTS: an autoregressive text-to-semantic model co-trained with ASR and a masked-generative semantic-to-acoustic model for synthesis only. End-to-end feedback across the text interface is enabled with straight-through argmax/Gumbel-Softmax and balanced with supervised ASR via dynamic weight averaging. Ablations examine optimal temperature schedules for in- and cross-domain transfer. Evaluation reveals TokenChain surpasses baseline accuracy 2-6 epochs earlier and yields 5-13% lower equal-epoch error with stable T2S on LibriSpeech, and reduces relative ASR WER by 56% and T2S WER by 31% on TED-LIUM with minimal forgetting, showing that chain learning remains effective with token interfaces and models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes TokenChain, a fully discrete speech chain that couples semantic-token ASR with a two-stage TTS (autoregressive text-to-semantic model co-trained with ASR plus masked-generative semantic-to-acoustic model). End-to-end feedback across the discrete interface is enabled via straight-through argmax (ASR-to-TTS) and Gumbel-Softmax (reverse), balanced by dynamic weight averaging with supervised ASR. Temperature-schedule ablations are performed for in- and cross-domain transfer. On LibriSpeech the method reaches baseline accuracy 2-6 epochs earlier with 5-13% lower equal-epoch error and stable T2S; on TED-LIUM it reports 56% relative ASR WER reduction and 31% T2S WER reduction with minimal forgetting.

Significance. If the gradient transmission across the semantic-token interface proves reliable, the work shows that classic machine speech-chain benefits can be realized with modern discrete token interfaces and models, offering a path to joint ASR-TTS improvement without continuous latent representations. The reported relative error reductions and faster convergence are practically relevant for low-resource or domain-adaptation settings, though the lack of error bars, full protocol details, and direct gradient diagnostics limits the strength of the evidence.

major comments (2)
  1. [Abstract and §4 (Results)] Abstract and §4 (Results): The central claims of 2-6 epochs earlier convergence, 5-13% lower equal-epoch error on LibriSpeech, and 56%/31% relative WER reductions on TED-LIUM rest on the assumption that straight-through argmax and Gumbel-Softmax transmit useful gradients across the semantic-token bottleneck. The manuscript provides no supporting diagnostics (gradient-norm statistics, mutual-information estimates across the interface, or detached-gradient control runs) to rule out that gains arise solely from the supervised ASR term or the two-stage TTS architecture rather than chain feedback.
  2. [§3.2 and §4.2] §3.2 and §4.2: Dynamic weight averaging and the temperature schedule are listed among the free parameters; the paper does not quantify how sensitive the headline improvements are to these choices or whether the reported temperature ablations were pre-specified versus post-hoc, which is required to establish that the chain-learning benefit is robust rather than tuned.
minor comments (2)
  1. [Abstract] Abstract: Concrete numerical claims (error reductions, epoch counts) are presented without error bars, standard deviations, or number of runs, which would improve interpretability of the reported gains.
  2. [§4.1] §4.1: The exact definition of 'stable T2S' and the precise baseline configurations (model sizes, training steps, data splits) should be stated explicitly to allow reproduction.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below and have revised the manuscript to strengthen the evidence for gradient transmission and hyperparameter robustness.

read point-by-point responses

  1. Referee: [Abstract and §4 (Results)] Abstract and §4 (Results): The central claims of 2-6 epochs earlier convergence, 5-13% lower equal-epoch error on LibriSpeech, and 56%/31% relative WER reductions on TED-LIUM rest on the assumption that straight-through argmax and Gumbel-Softmax transmit useful gradients across the semantic-token bottleneck. The manuscript provides no supporting diagnostics (gradient-norm statistics, mutual-information estimates across the interface, or detached-gradient control runs) to rule out that gains arise solely from the supervised ASR term or the two-stage TTS architecture rather than chain feedback.

    Authors: We agree that explicit diagnostics would make the contribution of chain feedback more convincing. In the revised manuscript we have added gradient-norm statistics across the semantic-token interface (new Figure 5) and a detached-gradient control ablation in §4.3. The control shows that blocking gradients through the token bottleneck eliminates the reported gains in convergence speed and WER, while the supervised ASR term alone cannot reproduce the full improvement. Mutual-information estimates were omitted due to prohibitive compute cost on our scale; we believe the direct control experiment provides stronger causal evidence than MI for the specific claim of useful gradient transmission. revision: yes

  2. Referee: [§3.2 and §4.2] §3.2 and §4.2: Dynamic weight averaging and the temperature schedule are listed among the free parameters; the paper does not quantify how sensitive the headline improvements are to these choices or whether the reported temperature ablations were pre-specified versus post-hoc, which is required to establish that the chain-learning benefit is robust rather than tuned.

    Authors: We have added a sensitivity study in the revised §4.2 that sweeps the dynamic-weight-averaging coefficient over [0.1, 0.9] and the temperature schedule over three families (linear, exponential, cosine). The headline improvements remain stable within ±2 % relative WER across this range. The temperature ablations were performed to characterize in- versus cross-domain transfer behavior as described in the original experimental design; we have now explicitly labeled them as such and included the full grid in the appendix to clarify they were not selected post-hoc for the main results. revision: partial

Circularity Check

0 steps flagged

No significant circularity: results are direct empirical evaluations on held-out sets.

full rationale

The paper describes TokenChain as a method coupling semantic-token ASR with a two-stage TTS, using straight-through argmax/Gumbel-Softmax for cross-interface gradients and dynamic weight averaging for balancing. Headline gains (earlier convergence by 2-6 epochs, WER reductions on LibriSpeech and TED-LIUM) are presented as outcomes of training and testing on standard benchmarks rather than algebraic identities or self-referential fits. No equations, parameter renamings, or self-citations are shown that collapse the reported metrics back to the training objectives by construction. The approach builds on established discrete-interface estimators without deriving its performance claims from the inputs themselves, rendering the derivation self-contained.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that semantic tokens form a sufficient discrete interface for both recognition and synthesis, plus several training hyperparameters whose values are selected or scheduled via ablation.

free parameters (2)
  • temperature schedule
    Optimal in-domain and cross-domain temperature schedules are determined through ablations and affect transfer performance.
  • dynamic weight averaging coefficients
    Weights balancing chain feedback against supervised ASR loss are adjusted during training.
axioms (1)
  • domain assumption Semantic tokens extracted from speech can serve as a lossless-enough interface for both ASR and subsequent TTS synthesis.
    The entire discrete chain is built on this representation choice.

pith-pipeline@v0.9.0 · 5696 in / 1456 out tokens · 40168 ms · 2026-05-18T09:12:25.566610+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 6 internal anchors

  1. [1]

    The machine speech chain operationalizes this by training automatic speech recogni- tion (ASR) and text-to-speech (TTS) in a closed loop [1]

    INTRODUCTION Human speech is a bidirectional mapping between sym- bolic text and acoustic realizations; coupling perception and production improves learning. The machine speech chain operationalizes this by training automatic speech recogni- tion (ASR) and text-to-speech (TTS) in a closed loop [1]. Prior work enabled backpropagation from TTS to ASR via st...

    work page

  2. [2]

    TokenChain: A Discrete Speech Chain via Semantic Token Modeling

    realizes semantic-to-acoustic generation in this manner, and MaskGCT [11] extends the approach to a two-stage text-to-semantic and semantic-to-acoustic framework. For ASR, scaling in self-supervised learning have driven gains, and discretizing SSL features emerged as viable inputs. Early unit-based systems required language models to match log-mel baselin...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Data Preparation Transcripts are tokenized with byte-pair encoding (BPE) [14] intoy= (y 1,

    METHODS 2.1. Data Preparation Transcripts are tokenized with byte-pair encoding (BPE) [14] intoy= (y 1, . . . , yL)from a vocabulary of sizeC. Speech is tokenized with SpeechTokenizer under semantic distilla- tion: RVQ-1 is guided toward the layerwise mean of HuBERT

    work page

  4. [4]

    We denote RVQ-1 codes as semantic tokenss= (s 1,

    to concentrate linguistic content, while RVQ-2:8 capture residual acoustic detail. We denote RVQ-1 codes as semantic tokenss= (s 1, . . . , sT )and higher-layer stacks as acoustic tokensa 2:8 witha j = (a j 1, . . . , aj T ). Each utterance is en- coded once;(s,y)pair is used in TokenChain training, and a short acoustic prompta p is retained only for audi...

    work page

  5. [5]

    Setup Datasets.We pretrain ASR/T2S on LibriSpeech-100 [21], and run chain training on LibriSpeech-960 and TED-LIUM v2 [22]

    EXPERIMENTS 3.1. Setup Datasets.We pretrain ASR/T2S on LibriSpeech-100 [21], and run chain training on LibriSpeech-960 and TED-LIUM v2 [22]. Audio synthesis uses a trained-and-frozen S2A on Emilia [23]. We evaluate post-chain ASR/T2S on Lib- riSpeech dev/test-{clean,other}and TED-LIUM dev/test. Framework.TokenChain framework is adopted from ESPnet

    work page

  6. [6]

    Metrics.For ASR, we evaluate CER and WER

    for ASR/chain training and Amphion [25] for T2S/S2A training/evaluation to form a unified TokenChain pipeline. Metrics.For ASR, we evaluate CER and WER. For syn- thesized audio, we evaluate WER via Whisper-large-v3 [26], speaker similarity (SIM-O) with WavLM–TDNN2 [27], and speech quality using UTMOSv2 (Predicted MOS) [28]. 3.2. Model and Training Details...

    work page 2048

  7. [7]

    CONCLUSIONS We introduced TokenChain, a machine speech chain with a fully discrete token interface. Using ST-argmax/Gumbel– Softmax with dynamic weight averaging, it enables end-to- end feedback between a semantic-token ASR and an AR T2S while keeping a NAR S2A fixed for synthesis. Empirically, TokenChain improves recognition under equal compute and conve...

    work page

  8. [8]

    ACKNOWLEDGMENTS This work was supported by the Guangdong Introducing In- novative and Entrepreneurial Teams Program

    work page

  9. [9]

    Listening while speaking: Speech chain by deep learning,

    A. Tjandra, S. Sakti, and S. Nakamura, “Listening while speaking: Speech chain by deep learning,” in2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2017, pp. 301–308

    work page 2017

  10. [10]

    End-to-end feedback loss in speech chain framework via straight-through estimator,

    A. Tjandra, S. Sakti, and S. Nakamura, “End-to-end feedback loss in speech chain framework via straight-through estimator,” in2019 IEEE Int. Conf. Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2019, pp. 6281–6285

    work page 2019

  11. [11]

    Exploring machine speech chain for domain adaptation and few-shot speaker adaptation,

    F. Yue, Y . Deng, L. He, and T. Ko, “Exploring machine speech chain for domain adaptation and few-shot speaker adaptation,” arXiv:2104.03815, 2021

    work page arXiv 2021

  12. [12]

    Recent advances in speech language models: A survey,

    W. Cui, D. Yu, X. Jiao, Z. Meng, G. Zhang, Q. Wang, Y . Guo, and I. King, “Recent advances in speech language models: A survey,”arXiv:2410.03751, 2024

    work page arXiv 2024

  13. [13]

    Recent advances in discrete speech tokens: A review

    Y . Guo, Z. Li, H. Wang, B. Li, C. Shao, H. Zhang, C. Du, X. Chen, S. Liu, and K. Yu, “Recent advances in discrete speech tokens: A review,”arXiv:2502.06490, 2025

    work page arXiv 2025

  14. [14]

    Soundstream: An end-to-end neural audio codec,

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “Soundstream: An end-to-end neural audio codec,”IEEE/ACM Trans. Audio, Speech, and Language Pro- cessing (TASLP), vol. 30, pp. 495–507, 2021

    work page 2021

  15. [15]

    Speechtokenizer: Unified speech tokenizer for speech large language models

    X. Zhang, D. Zhang, S. Li, Y . Zhou, and X. Qiu, “Speechtok- enizer: Unified speech tokenizer for speech large language mod- els,”arXiv:2308.16692, 2023

    work page arXiv 2023

  16. [16]

    Towards con- trollable speech synthesis in the era of large language models: A survey,

    T. Xie, Y . Rong, P. Zhang, W. Wang, and L. Liu, “Towards con- trollable speech synthesis in the era of large language models: A survey,”arXiv:2412.06602, 2024

    work page arXiv 2024

  17. [17]

    Audiolm: a language modeling approach to audio generation,

    Z. Borsos, R. Marinier, D. Vincent, E. Kharitonov, O. Pietquin, M. Sharifi, D. Roblek, O. Teboul, D. Grangier, M. Tagliasac- chi, et al., “Audiolm: a language modeling approach to audio generation,”IEEE/ACM Trans. Audio, Speech, and Language Processing (TASLP), vol. 31, pp. 2523–2533, 2023

    work page 2023

  18. [18]

    Soundstorm: Efficient parallel audio generation,

    Z. Borsos, M. Sharifi, D. Vincent, E. Kharitonov, N. Zeghidour, and M. Tagliasacchi, “Soundstorm: Efficient parallel audio gen- eration,”arXiv:2305.09636, 2023

    work page arXiv 2023

  19. [19]

    Maskgct: Zero- shot text-to-speech with masked generative codec transformer,

    Y . Wang, H. Zhan, L. Liu, R. Zeng, H. Guo, J. Zheng, Q. Zhang, X. Zhang, S. Zhang, and Z. Wu, “Maskgct: Zero- shot text-to-speech with masked generative codec transformer,” arXiv:2409.00750, 2024

    work page arXiv 2024

  20. [20]

    Effectiveness of self-supervised pre-training for asr,

    A. Baevski and A. Mohamed, “Effectiveness of self-supervised pre-training for asr,” in2020 IEEE Int. Conf. Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 7694–7698

    work page 2020

  21. [21]

    Ex- ploration of efficient end-to-end asr using discretized input from self-supervised learning,

    X. Chang, B. Yan, Y . Fujita, T. Maekaku, and S. Watanabe, “Ex- ploration of efficient end-to-end asr using discretized input from self-supervised learning,”arXiv:2305.18108, 2023

    work page arXiv 2023

  22. [22]

    Neural Machine Translation of Rare Words with Subword Units

    R. Sennrich, B. Haddow, and A. Birch, “Neural machine trans- lation of rare words with subword units,”arXiv:1508.07909, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  23. [23]

    Hubert: Self-supervised speech representa- tion learning by masked prediction of hidden units,

    W. Hsu, B. Bolte, Y . H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hubert: Self-supervised speech representa- tion learning by masked prediction of hidden units,”IEEE/ACM Trans. Audio, Speech, and Language Processing (TASLP), vol. 29, pp. 3451–3460, 2021

    work page 2021

  24. [24]

    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

    work page 2017

  25. [25]

    Llama 2: Open Foundation and Fine-Tuned Chat Models

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al., “Llama 2: Open foundation and fine-tuned chat models,” arXiv:2307.09288, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

    Y . Bengio, N. L´eonard, and A. Courville, “Estimating or prop- agating gradients through stochastic neurons for conditional computation,”arXiv:1308.3432, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  27. [27]

    Categorical Reparameterization with Gumbel-Softmax

    E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,”arXiv:1611.01144, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  28. [28]

    End-to-end multi-task learning with attention,

    S. Liu, E. Johns, and A. J. Davison, “End-to-end multi-task learning with attention,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2019, pp. 1871–1880

    work page 2019

  29. [29]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in2015 IEEE Int. Conf. Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2015, pp. 5206–5210

    work page 2015

  30. [30]

    Enhancing the ted- lium corpus with selected data for language modeling and more ted talks.,

    A. Rousseau, P. Del´eglise, Y . Esteve, et al., “Enhancing the ted- lium corpus with selected data for language modeling and more ted talks.,” inLREC, 2014, pp. 3935–3939

    work page 2014

  31. [31]

    Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,

    H. He, Z. Shang, C. Wang, X. Li, Y . Gu, H. Hua, L. Liu, C. Yang, J. Li, P. Shi, et al., “Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 885–890

    work page 2024

  32. [32]

    ESPnet: End-to-End Speech Processing Toolkit

    S. Watanabe, T. Hori, S. Karita, T. Hayashi, J. Nishitoba, Y . Unno, N. E. Y . Soplin, J. Heymann, M. Wiesner, N. Chen, et al., “Espnet: End-to-end speech processing toolkit,” arXiv:1804.00015, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  33. [33]

    Amphion: an open-source audio, music, and speech generation toolkit,

    X. Zhang, L. Xue, Y . Gu, Y . Wang, J. Li, H. He, C. Wang, S. Liu, X. Chen, J. Zhang, et al., “Amphion: an open-source audio, music, and speech generation toolkit,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 879– 884

    work page 2024

  34. [34]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInt. Conf. Machine Learning (ICML). PMLR, 2023, pp. 28492–28518

    work page 2023

  35. [35]

    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. Xiao, 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

    work page 2022

  36. [36]

    The t05 sys- tem for the V oiceMOS Challenge 2024: Transfer learning from deep image classifier to naturalness MOS prediction of high- quality synthetic speech,

    K. Baba, W. Nakata, Y . Saito, and H. Saruwatari, “The t05 sys- tem for the V oiceMOS Challenge 2024: Transfer learning from deep image classifier to naturalness MOS prediction of high- quality synthetic speech,” inIEEE Spoken Language Technol- ogy Workshop (SLT), 2024

    work page 2024