Pith. sign in

REVIEW 4 major objections 6 minor 30 references

Analytic Study of Text-Free Speech Synthesis for Raw Audio using a Self-Supervised Learning Model

T0 review · 4 major / 6 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 Using discrete self-supervised speech tokens as the input to Tacotron 2 preserves prosody and acoustic quality better than using ASR transcripts, at the cost of some intelligibility.

desk verdict Useful empirical map of SSL-token vs ASR-text TTS, but the prosody-preservation claim is untested because WARP-Q/SDR use GT-synthesized speech as reference, not the original audio. read the letter →

arxiv 2412.03074 v1 pith:FVIFMNPG submitted 2024-12-04 cs.CL cs.SDeess.AS

Joonyong Park , Daisuke Saito , Nobuaki Minematsu This is my paper · ORCID
classification cs.CLcs.SDeess.AS
keywords speechsynthesisself-supervisedlearningdiscreterepresentationstext-freezero-resourceTacotron2HuBERTprosody
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 speech synthesis can be driven directly by raw audio, without any transcribed text, using discrete tokens extracted by a self-supervised speech model. It builds three systems with identical Tacotron 2 backbones and varying input: ground-truth text, transcripts produced by an ASR model, and discrete symbol sequences from a self-supervised model (HuBERT features quantized with k-means). Across English and Japanese, the paper reports that ASR text better preserves semantic and phonetic content, while the self-supervised discrete tokens better preserve prosodic, intonational, and general acoustic quality, as measured by UTMOS, WARP-Q, and SDR. The finding matters because it suggests a route to zero-resource, multilingual speech synthesis that needs no text annotations.

What carries the argument

The load-bearing object is the discrete symbol sequence produced by a self-supervised speech model: raw audio is fed into HuBERT, features from a chosen Transformer layer (6th or 12th) are clustered with k-means into a codebook of 50, 200, or 1000 symbols, and the resulting token stream is used in place of text as input to Tacotron 2. This representation is what isolates the comparison: the synthesis network, vocoder, and training data are held fixed, so any difference between the SSL-token system and the ASR-text system is attributable to the information content of the input representation. The paper also varies whether the SSL encoder was trained on the same language as the synthesizer and whether repeated tokens are collapsed, and reports that larger codebooks, later layers, and language-matched encoders improve metrics.

What would settle it

Run the Japanese condition again with a high-accuracy ASR (character error rate below 5% on the test set) and compare the same SSL configurations against its transcripts; if the SSL advantage in UTMOS, WARP-Q, and SDR shrinks to within confidence intervals, the paper's claim that SSL tokens preserve acoustic content better than text would not hold for that language.

Watch

Extended reading notes

Core claim

Using discrete symbol representations from the SSL model instead of ASR-derived text as input to Tacotron 2 yields higher UTMOS, WARP-Q, and SDR scores, and slightly lower intelligibility metrics. In the English condition, the best SSL configuration (language-matched, 1000-cluster codebook, 12th HuBERT layer) reaches UTMOS 3.72, WARP-Q 2.93, and SDR -17.05 dB, compared with UTMOS 3.41, WARP-Q 2.71, and SDR -18.55 dB for the ASR baseline, while the ASR baseline has lower error rates in WER, PER, and CER. The paper concludes that text representations carry an advantage for semantic content, whereas discrete SSL representations are superior for acoustic content including prosody and intonation. The effect holds for both English and Japanese, with Japanese showing smaller SSL advantages and a weaker overall intelligibility level.

Load-bearing premise

The load-bearing premise is that ASR transcripts are a representative text baseline, but the ASR model errs on roughly one in five Japanese characters, so the text condition is a noisy proxy for true text input.

Editorial extensions

If this is right

  • Speech synthesis from untranscribed audio is feasible with only small quality losses relative to text-driven synthesis, and with gains in naturalness and acoustics over ASR-derived text.
  • Larger discrete codebooks improve both intelligibility and naturalness, with the largest jump occurring between 50 and 200 tokens, suggesting a point of diminishing returns.
  • The SSL encoder's training language matters: language-matched speech2unit/unit2speech pairs produce more intelligible and more natural synthetic speech than cross-language pairs.
  • Choosing a deeper Transformer layer (12th vs 6th) yields more semantic content, so layer selection is a practical control knob for the linguistic/acoustic trade-off in SSL-token synthesis.
  • ASR-based text remains the better input when intelligibility is the priority, so the two representation families are complementary rather than one universally dominant.

Reading between the lines

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

  • A direct testable extension would feed both ASR text and SSL tokens into the synthesizer; if the complementarity is real, the combined input should raise both intelligibility and acoustic quality above either alone.
  • Because the Japanese ASR baseline itself has a 19.35% character error rate on ground-truth test audio, the reported SSL advantage in Japanese may be partly an artifact of a weak text baseline; repeating the comparison with a stronger recognizer would show whether the acoustic advantage survives.
  • The quality metrics WARP-Q and SDR compare SSL/ASR outputs against synthesis from ground-truth scripts as reference; thus the 'acoustic superiority' of SSL tokens is relative to the reference generation pipeline, and a different reference could shift the ranking.
  • If SSL tokens encode prosodic contours well, the finding suggests using untranscribed audiobook or conversational data to build expressive synthesis systems in low-resource languages without transcription campaigns.
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

4 major / 6 minor

Summary. The paper studies whether discrete symbol representations extracted from a self-supervised learning (SSL) model (HuBERT with k-means quantization) can replace conventional text as input to a Tacotron 2 speech synthesizer, in a text-free speech synthesis setup. The authors build three input conditions—ground-truth text (GT), Whisper ASR transcripts (ASR), and SSL discrete units—for English and Japanese. They evaluate synthesized speech on intelligibility (WER/CER/PER), naturalness (UTMOS), and acoustic quality (WARP-Q, SDR), varying the SSL codebook size (50/200/1000), HuBERT layer (6/12), and language match between the SSL encoder and the synthesizer. The central claim, stated in the abstract and conclusion, is that SSL discrete units preserve acoustic content, including prosodic and intonational information, better than ASR text, whereas ASR text better preserves semantic/linguistic content.

Significance. If the central claim were fully supported, the paper would be a useful empirical reference for zero-resource and textless speech synthesis. Its main strengths are the systematic ablation across codebook sizes, Transformer layers, and language-matched/unmatched conditions, judged on a common benchmark with publicly available components. The rich results table (Table II) provides a useful comparison of many configurations. However, the headline claim about acoustic content preservation, specifically prosody and intonation, rests on metrics whose references do not actually measure preservation of the original utterance's acoustic properties, as detailed in the major comments. The study is therefore a useful data point, but its main conclusion is currently overstated.

major comments (4)
  1. [§IV-C, Table II] The WARP-Q evaluation uses synthesized speech from correct scripts as the reference, not the original test audio, as stated in §IV-C: 'synthesized speech using correct scripts was used as a reference.' Therefore, a higher WARP-Q value for an SSL condition means that the SSL output is closer to a text-driven canonical rendering of the same sentence, not that it preserves the original utterance's prosody, intonation, or other acoustic content. The abstract's claim that SSL discrete symbols are 'superior for preserving acoustic content, including prosodic and intonational information' is not supported by the WARP-Q results, and the conclusion repeats this overstatement. Use the original source audio as the reference if acoustic-content preservation is the intended claim, or explicitly reframe the metric as 'closeness to text-conditioned synthesis.'
  2. [§IV-C, Table II] The SDR evaluation is described as measuring degradation 'relative to the input audio,' but the reference signal is never defined in the text. Table II reports SDR only for ASR and SSL conditions, with no SDR value for the GT condition, so the reference cannot be the GT-text-synthesized speech (which would make the GT SDR trivially 0). If the reference is the original test audio, the paper must state this explicitly and explain how the temporal alignment between the synthesized speech and the original audio is handled, given that Tacotron 2 outputs are not time-aligned to the source. As written, the SDR column is uninterpretable, and the conclusion that SSL output is less noisy than ASR output is not substantiated.
  3. [§IV-A and Table II] No significance tests or confidence intervals are provided for the WER, CER, and PER results, although several decisive comparisons rely on very small differences. For example, in English, the ASR condition has WER 5.41% and the best SSL condition (match-1000-L12) has WER 5.45%; these are effectively identical. Similarly, the UTMOS differences between ASR and the best SSL conditions are around 0.1–0.3 points, and the quoted confidence intervals in footnote 2 are explicitly derived from the model output standard deviation, not from human ratings, so they do not support claims of statistically significant naturalness differences. The paper should report error bars or statistical tests for all intelligibility and naturalness metrics, or weaken the claims accordingly.
  4. [§IV-A, Table II] The paper reports that Whisper-base has a Character Error Rate of 19.35% on the Japanese GT test audio. This makes the ASR condition in Japanese a degraded label rather than a reliable text representation. Consequently, the SSL-versus-ASR comparisons in Japanese, which contribute to the conclusion that SSL representations are superior for acoustic content, compare against an unusually weak baseline. The authors should either use a stronger ASR model for the Japanese condition, report the acoustic comparisons separately for English and Japanese while acknowledging the baseline quality, or temper the cross-language generality of the conclusion.
minor comments (6)
  1. [Abstract] The phrase 'the proposed speech synthesis is conducted using discrete symbol representations' is misleading because no new synthesis model is proposed; the paper is a comparative study. Consider rewording to 'we conduct speech synthesis using discrete symbol representations.'
  2. [Throughout] The typesetting of 'W ARP-Q' and 'V ocoder' (with spaces) appears to be a LaTeX artifact; these should read 'WARP-Q' and 'Vocoder' throughout.
  3. [§III-A] The Japanese morphological analyzer is referred to as 'Mecab'; the standard name is 'MeCab'.
  4. [§IV-A] In the first paragraph, it is not immediately clear whether WER/CER are computed by feeding the synthesized speech into Whisper and comparing its output to the reference transcript, or by comparing Whisper's output on the input transcripts. Please clarify the exact evaluation procedure.
  5. [§IV-B] The sentence 'while some errors are within confidence levels' is vague; specify which comparisons are within the confidence intervals and which are not.
  6. [Table II] For the GT rows, WARP-Q and SDR are listed as '−'. The caption should note that these metrics are undefined for GT because GT speech serves as the reference condition (or state the actual reason).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the comparison is self-contained, externally benchmarked, and does not reduce its conclusion to its inputs.

full rationale

The paper is a controlled empirical comparison, not a derivation: three input representations (ground-truth text, Whisper ASR transcripts, and HuBERT/k-means discrete units) feed the same Tacotron 2/vocoder backbone, and the reported metrics are external benchmarks rather than quantities fitted to the hypothesis. Intelligibility is measured against human-written scripts (WER/CER) and via the independent allosaurus phoneme recognizer (PER); naturalness uses the pretrained UTMOS predictor; audio quality uses WARP-Q and SDR, with SDR described as measuring the output relative to the input audio. No parameter is fitted to the test metrics, and no load-bearing self-citation appears in the argument; the cited prior work is by other research groups. Two evaluation-design caveats exist but are not circularity: WARP-Q uses synthesized speech from correct scripts as its reference rather than the original test audio, and Whisper is used both to create the ASR condition and to compute WER/CER on all conditions. These concerns affect construct validity and could inflate or deflate measured differences, but they do not make the SSL-versus-ASR comparison equivalent to its inputs by construction. The central claim therefore remains an externally checkable empirical result, so the circularity score is 0.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central claims rest on several domain assumptions about the transferability of pretrained models to Japanese and the validity of automatic metrics, but the paper introduces no fitted parameters or invented entities. Experimental conditions such as codebook sizes and layers are ablations, not free parameters.

assumptions (6)
  • domain assumption HuBERT-base pretrained on English LibriSpeech provides discrete units that transfer adequately to Japanese when combined with a Japanese k-means codebook.
    Section III-A uses HuBERT-base for both English and Japanese speech2unit modules, relying on cross-lingual transfer.
  • domain assumption Whisper-base is an acceptable ASR for both English and Japanese for creating baseline text labels.
    Section III-A specifies Whisper-base for both languages, despite the paper's own Japanese CER of 19.35% on GT speech.
  • domain assumption UTMOS is a valid automatic proxy for human-rated naturalness.
    Section IV-B uses UTMOS as the naturalness metric; the paper notes in footnote 2 that it is an objective model output, not a subjective MOS.
  • domain assumption WARP-Q and SDR are valid metrics for comparing resynthesized speech quality when referenced against GT-synthesized speech.
    Section IV-C applies these codec-oriented metrics to synthesized speech, referencing the GT-synthesized output.
  • domain assumption The same Tacotron 2 architecture and vocoder across all conditions gives a fair comparison of input representations.
    Section III-A states that Tacotron 2 is used under identical conditions except for input representation.
  • domain assumption Removing consecutive repeated symbols with threshold n=1 does not discard essential prosodic information.
    Section III-A fixes n=1 for all experiments, which could affect rhythm and duration cues in the synthesized speech.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analytic Study of Text-Free Speech Synthesis for Raw Audio using a Self-Supervised Learning Model." pith.science (2026). https://pith.science/paper/FVIFMNPG

@misc{pith2026241203074,
  author       = {Pith},
  title        = {Pith review of: Analytic Study of Text-Free Speech Synthesis for Raw Audio using a Self-Supervised Learning Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FVIFMNPG}},
  note         = {Machine review of arXiv:2412.03074}
}
read the original abstract

We examine the text-free speech representations of raw audio obtained from a self-supervised learning (SSL) model by analyzing the synthesized speech using the SSL representations instead of conventional text representations. Since raw audio does not have paired speech representations as transcribed texts do, obtaining speech representations from unpaired speech is crucial for augmenting available datasets for speech synthesis. Specifically, the proposed speech synthesis is conducted using discrete symbol representations from the SSL model in comparison with text representations, and analytical examinations of the synthesized speech have been carried out. The results empirically show that using text representations is advantageous for preserving semantic information, while using discrete symbol representations is superior for preserving acoustic content, including prosodic and intonational information.

Figures

Figures reproduced from arXiv: 2412.03074 by the authors.

Figure 1
Figure 1. (a) Architecture of GSLM and (b) Application to Japanese Language [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Building a speech synthesis system using (a) ground-truth script [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. An example of a SSL system that speech2unit-unit2speech pairs are [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 17 canonical work pages

  1. [1]

    A brief overview of unsupervised neural speech representation learning,

    L. Borgholt, J. D. Havtorn, J. Edin, L. Maaløe, and C. Igel, “A brief overview of unsupervised neural speech representation learning,” CoRR, vol. abs/2203.01829,

    arXiv

  2. [2]

    J. Shen, R. Pang, R. J. Weiss, et al. , Natural tts synthesis by conditioning wavenet on mel spectrogram predictions, 2018. arXiv: 1712.05884

    arXiv 2018

  3. [3]

    Fastspeech 2: Fast and high-quality end-to-end text to speech,

    Y . Ren, C. Hu, X. Tan, et al. , “Fastspeech 2: Fast and high-quality end-to-end text to speech,” ArXiv, vol. abs/2006.04558, 2020

    arXiv 2006

  4. [4]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” in Proceedings of the 38th Interna- tional Conference on Machine Learning , vol. 139, 2021, pp. 5530–5540. 5

    work page 2021

  5. [5]

    Natural TTS synthesis by conditioning Wavenet on mel-spectrogram predictions,

    J. Shen, R. Pang, R. J. Weiss, et al. , “Natural TTS synthesis by conditioning Wavenet on mel-spectrogram predictions,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing , 2018, pp. 4779–4783

    2018

  6. [6]

    On generative spoken language modeling from raw audio,

    K. Lakhotia, E. Kharitonov, W.-N. Hsu, et al. , “On generative spoken language modeling from raw audio,” Transactions of the Association for Computational Lin- guistics, vol. 9, pp. 1336–1354, 2021

    2021

  7. [7]

    Representa- tion learning with contrastive predictive coding,

    A. van den Oord, Y . Li, and O. Vinyals, “Representa- tion learning with contrastive predictive coding,” CoRR, vol. abs/1807.03748, 2018. arXiv: 1807.03748

    arXiv 2018

  8. [8]

    Wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “Wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Advances in Neural Infor- mation Processing Systems , vol. 33, 2020, pp. 12 449– 12 460

    2020

Show all 30 references
  1. [9]

    HuBERT: Self- supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self- supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 29, pp. 3451–3460, 2021

    2021

  2. [10]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, et al. , “Attention is all you need,” in Advances in Neural Information Processing Systems , vol. 30, Curran Associates, Inc., 2017

    2017

  3. [11]

    The Zero Resource Speech Challenge 2020: Discovering Discrete Subword and Word Units,

    E. Dunbar, J. Karadayi, M. Bernard, et al. , “The Zero Resource Speech Challenge 2020: Discovering Discrete Subword and Word Units,” in Proc. Interspeech 2020 , 2020, pp. 4831–4835

    2020

  4. [12]

    The zero resource speech challenge 2021: Spoken language mod- elling,

    E. Dunbar, M. Bernard, N. Hamilakis, et al., “The zero resource speech challenge 2021: Spoken language mod- elling,” in Proc. Interspeech 2021, 2021, pp. 1574–1578

    2021

  5. [13]

    Self- supervised language learning from raw audio: Lessons from the zero resource speech challenge,

    E. Dunbar, N. Hamilakis, and E. Dupoux, “Self- supervised language learning from raw audio: Lessons from the zero resource speech challenge,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1211–1226, 2022

    2022

  6. [14]

    A comparison of discrete and soft speech units for improved voice conversion,

    B. van Niekerk, M.-A. Carbonneau, J. Za ¨ıdi, M. Baas, H. Seut ´e, and H. Kamper, “A comparison of discrete and soft speech units for improved voice conversion,” ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 6562–6566, 2021

    2022

  7. [15]

    Textless direct speech-to- speech translation with discrete speech representation,

    X. Li, Y . Jia, and C.-C. Chiu, “Textless direct speech-to- speech translation with discrete speech representation,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2023, pp. 1–5. DOI: 10 . 1109 / ICASSP49357 . 2023 . 10096797

    2023

  8. [16]

    UnitSpeech: Speaker-adaptive Speech Synthesis with Untranscribed Data,

    H. Kim, S. Kim, J. Yeom, and S. Yoon, “UnitSpeech: Speaker-adaptive Speech Synthesis with Untranscribed Data,” in Proc. INTERSPEECH 2023 , 2023, pp. 3038–

    2023

  9. [17]

    LibriSpeech: An ASR corpus based on public domain audio books,

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

    2015

  10. [18]

    Ito and L

    K. Ito and L. Johnson, The lj speech dataset , https : //keithito.com/LJ-Speech-Dataset/, 2017

    2017

  11. [19]

    Y . Yue, M. Daijiro, and F. Seiji, Reazonspeech: A free and massive corpus for Japanese ASR , https://research. reazon.jp/ static/reazonspeech nlp2023.pdf, 2023

    2023

  12. [20]

    JSUT corpus: Free large-scale Japanese speech corpus for end- to-end speech synthesis,

    R. Sonobe, S. Takamichi, and H. Saruwatari, “JSUT corpus: Free large-scale Japanese speech corpus for end- to-end speech synthesis,” CoRR, vol. abs/1711.00354,

    arXiv

  13. [21]

    JVS corpus: Free Japanese multi- speaker voice corpus,

    S. Takamichi, K. Mitsui, Y . Saito, T. Koriyama, N. Tanji, and H. Saruwatari, “JVS corpus: Free Japanese multi- speaker voice corpus,” CoRR, vol. abs/1908.06248,

    1908 arXiv

  14. [22]

    Audio- book speech synthesis conditioned by cross-sentence context-aware word embeddings,

    W. Nakata, T. Koriyama, S. Takamichi, et al. , “Audio- book speech synthesis conditioned by cross-sentence context-aware word embeddings,” in Proc. The 11th ISCA SSW , 2021

    2021

  15. [23]

    J-MAC: Japanese multi-speaker audiobook corpus for speech synthesis,

    S. Takamichi, N. Wataru, T. Naoko, and S. Hiroshi, “J-MAC: Japanese multi-speaker audiobook corpus for speech synthesis,” in Interspeech 2022, 23rd Annual Conference of the International Speech Communication Association, ISCA, 2022, pp. 2358–2362

    2022

  16. [24]

    Radford, K

    A. Radford, K. Jong Wook, X. Tao, B. Greg, M. Chris- tine, and S. Ilya, Robust speech recognition via large- scale weak supervision. https://cdn.openai.com/papers/ whisper.pdf, 2022

    2022

  17. [25]

    Investi- gation of robustness of hubert features from different layers to domain, accent and language variations,

    P. Kumar, V . N. Sukhadia, and S. Umesh, “Investi- gation of robustness of hubert features from different layers to domain, accent and language variations,” in ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2022, pp. 6887–6891

    2022

  18. [26]

    ContentVec: An improved self-supervised speech representation by dis- entangling speakers,

    K. Qian, Y . Zhang, H. Gao, et al. , “ContentVec: An improved self-supervised speech representation by dis- entangling speakers,” in Proceedings of the 39th In- ternational Conference on Machine Learning , 2022, pp. 18 003–18 017

    2022

  19. [27]

    JSSS: free japanese speech corpus for summarization and simplification,

    S. Takamichi, M. Komachi, N. Tanji, and H. Saruwatari, “JSSS: free japanese speech corpus for summarization and simplification,” CoRR, vol. abs/2010.01793, 2020. arXiv: 2010.01793

    2010 arXiv

  20. [28]

    Universal phone recognition with a multilingual allophone system,

    X. Li, S. Dalmia, J. Li, et al. , “Universal phone recognition with a multilingual allophone system,” in ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2020, pp. 8249–8253

    2020

  21. [29]

    Speech quality assessment with W ARP-Q: From sim- ilarity to subsequence dynamic time warp cost,

    W. A. Jassim, J. Skoglund, M. Chinen, and A. Hines, “Speech quality assessment with W ARP-Q: From sim- ilarity to subsequence dynamic time warp cost,” IET Signal Processing, vol. 16, no. 9, 1050––1070, 2022. 6

    2022

  22. [3042]

    DOI: 10.21437/Interspeech.2023-2326

    2023 doi

Pith tools

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