Pith. sign in

REVIEW 4 major objections 4 minor 45 references

LibriTTS-VI: A Public Corpus and Novel Methods for Efficient Voice Impression Control

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read This paper claims that text-to-speech can take an 11-number voice-impression vector as the only control signal, without a reference recording, and follow it more faithfully than prior reference-based methods.

desk verdict Useful corpus and a sensible leakage-reduction trick, but the central controllability numbers are mostly self-referential; the subjective check is too thin to fully break the loop. read the letter →

arxiv 2509.15626 v4 pith:LPOIODXA submitted 2025-09-19 cs.SD eess.AS

classification cs.SDeess.AS
keywords voiceimpressioncontroltext-to-speechzero-shotTTSreference-freesynthesisdisentanglementleakagecorpusVITS
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

Voice impression control lets users adjust qualities like brightness or calmness numerically in synthesized speech, but prior systems needed a reference audio clip whose own impression leaked into the output. This paper introduces the first public corpus for this task, LibriTTS-VI, and proposes two ways to reduce leakage: training with separate utterances for speaker identity versus impression, and a reference-free model that generates the speaker embedding from the target impression vector alone. The paper reports that the reference-free method lowers the mean-squared error of 11-dimensional impression vectors from 0.61 to 0.41 objectively and from 1.15 to 0.92 subjectively, while keeping audio quality about the same. If right, fine-grained and privacy-friendly voice control becomes feasible without any reference voice.

What carries the argument

The control module inserted into the VITS-based TTS backbone. VIC-base fuses the target VI vector with the reference audio's encoded vector, using dropout and a gradient reversal layer to suppress impression information. VIC-sep feeds the module a second utterance from the same speaker as the speaker-identity source while the original utterance supplies target VI and synthesis text. VIC-rfg removes the reference: the module gets only the VI vector plus a random Gaussian vector, and the stochastic duration predictor is fine-tuned jointly to keep speaking rate stable.

What would settle it

Have new, independent annotators rate outputs of VIC-rfg versus VIC-base on a fresh set of utterances and targets, with annotators blind to which system generated each sample and with the VIE entirely out of the loop. If the reference-free model's perceived accuracy is no better than baseline (or no better than chance), the claim that reference-free conditioning improves controllability would be refuted.

Watch

Extended reading notes

Core claim

Central claim: impression leakage—the unwanted bias of the synthesized voice toward the reference audio's impression—can be largely removed by architecture. VIC-rfg, the reference-free system, replaces the reference audio's encoded vector with Gaussian noise inside the speaker-encoder control module, leaving the 11-dimensional target VI vector as the only speaker-related conditioning. In zero-shot evaluation on 39 unseen speakers, this reduces the leakage gap ΔV from 0.22 to 0.05, improves the average control slope from 0.096 to 0.177, and lowers subjective multi-VI MSE from 1.15 to 0.92. The paper also releases LibriTTS-VI, a public corpus of manual and estimated 11-dimensional impression l

Load-bearing premise

The central results rest on the Voice Impression Estimator (VIE) being a reliable proxy for human-perceived voice impression, because the same model creates the target labels during training and computes the objective error used to report success.

Editorial extensions

If this is right

  • VIC-rfg synthesizes speech from a target VI vector alone, so a user can request e.g. a brighter or calmer voice numerically without a reference recording.
  • LibriTTS-VI, with manual annotations and estimated VIs over LibriTTS-R, becomes a public benchmark for comparing VI control systems.
  • RVI-MSE and ΔV quantify leakage; the paper reports ΔV dropping from 0.22 (baseline) to 0.14 (VIC-sep) and 0.05 (VIC-rfg).
  • Control remains partial: the average modulation slope rises to 0.177 for VIC-rfg but is near zero for Powerful-Weak, so some dimensions still respond weakly.
  • Reference-free control trades speaker similarity: SECS falls from 0.82–0.84 to 0.76, a quality to weigh in applications needing voice preservation.

Reading between the lines

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

  • Inference: Since the same VIE supplies training labels and objective metrics and was built from only 100 annotated utterances with average inter-annotator agreement α=0.464, the reported MSE gains may partially reflect fitting to the estimator; a blind listening study with fresh raters would test whether perceived controllability improves as much as the numbers say.
  • Inference: The reference-free architecture opens an interface where users adjust a voice's impression via sliders or a text-to-vector mapping without ever providing a reference recording; the paper does not evaluate such an interface.
  • Inference: The RVI-MSE and ΔV protocol could serve as a general leakage test for any conditioning attribute in controllable TTS, not just voice impression; that broader use is not explored here.
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

4 major / 4 minor

Summary. The paper introduces LibriTTS-VI, a public corpus of 11-dimensional voice impression (VI) annotations built on LibriTTS-R, and proposes two methods to reduce impression leakage in zero-shot TTS: VIC-sep, which uses separate utterances for speaker and VI conditioning, and VIC-rfg, which removes the reference audio entirely and generates a speaker embedding from the target VI. The authors report objective and subjective improvements in controllability (RVI-MSE from 0.61 to 0.41; multiple-VI subjective MSE from 1.15 to 0.92) while maintaining synthesis quality.

Significance. If the evaluation is valid, the paper makes a useful contribution: LibriTTS-VI would be the first public VI corpus with explicit annotation standards, and the reference-free VIC-rfg architecture is a practical step toward privacy-preserving and leakage-free VI control. The authors also provide reproducible experimental details and release the corpus. However, the central quantitative claims currently rest on a partially self-referential evaluation loop and a small, non-independent subjective test, so the evidence for the headline improvements is not yet convincing.

major comments (4)
  1. [Sec. 4.2, Eqs. (6)-(8); Sec. 2.2, Eq. (5); Table 3] The objective controllability metrics VI-MSE, RVI-MSE, and ΔV are computed with the same VIE that supplies target VI vectors during fine-tuning. The VIE is trained on 100 manually annotated utterances with average Krippendorff's α = 0.464 (Table 1) and labels propagated by acoustic similarity without validation. The reported 0.61→0.41 improvement therefore shows alignment with this particular estimator, not necessarily with human perception. An independent perceptual test, or at least a VIE validated on held-out human ratings, is needed to support the headline claim.
  2. [Sec. 4.3, Table 5] The subjective controllability evaluation uses the same four annotators who created the training labels and reports only speaker 8555. This does not break the circularity: the annotators may be rating consistently with their own earlier labels, and a single speaker is not representative. The MOS test with 30 independent raters is a strength, but the controllability claim needs more speakers and independent (non-author) raters.
  3. [Sec. 4.2, Eq. (8); Sec. 2.4; Table 3] For VIC-rfg, the reference audio r is not used at all by construction, so ΔV = RVI-MSE - VI-MSE does not measure leakage from a reference; it only compares the model's error for two different target vectors. The near-zero ΔV (0.05) is therefore largely an architectural consequence, not an empirical demonstration of leakage reduction. The paper should state this explicitly and not present ΔV for VIC-rfg as comparable to the baseline's leakage measure.
  4. [Abstract, last sentence] The abstract claims that a comparison with a prompt-based TTS reveals imprecise numerical control and VI-text entanglement that the proposed methods overcome. No such experiment or result appears in the manuscript. Either add the missing comparison or remove this claim from the abstract.
minor comments (4)
  1. [Table 1] K) Slow–Fast is listed with α = '-' because it is derived objectively, but the reported average α = 0.464 appears to be over the other 10 dimensions. Please clarify in the table caption or footnote.
  2. [Sec. 4.2, modulation experiment] The target VI is modulated from -3 to +3, but the annotation scale is 1–7. Specify how the anchor vector and modulation range are mapped to the 1–7 scale.
  3. [Sec. 4.3] The text says 'the same audio generated' from the modulation experiment (50 sentences per condition) is used, but Table 5 reports 10 sentences per condition. Clarify the selection or the number of sentences.
  4. [Eqs. (6)-(8)] VIC-rfg uses a sampled Gaussian noise vector, so synthesis is stochastic; Eq. (8) implicitly assumes determinism. State whether the metrics are averaged over noise samples.

Circularity Check

2 steps flagged · score 6.0 of 10

Objective gains rest on a closed evaluation loop: the VIE supplies the training targets and also computes the reported metrics; VIC-rfg's low ΔV is an architectural tautology.

  1. fitted input called prediction [Sec. 2.2 (Eqs. 4-5) and Sec. 4.2 (Eqs. 6-8), Table 3]
    "During the fine-tuning, the target VI vector v is supplied by a pre-trained voice impression estimator (VIE) ... RVI-MSE = ||v∗ − VIE(s(r, v∗))||2 2"

    The VIE is trained on manual labels and then its outputs become the target VI vectors used to fine-tune the control module. The objective controllability metrics VI-MSE and RVI-MSE are computed with the same VIE: the target v/v* are VIE estimates and the error is the difference between those estimates and VIE's prediction on synthesized audio. The reported improvement (0.61→0.41) therefore measures how well the TTS model matches the very estimator that generated its training targets; it is a self-consistency score, not an independent perceptual measurement. Since the VIE itself is trained on only 100 annotations with average Krippendorff α=0.464, the shared bias can be carried into both the training target and the evaluation, so the objective gain may reflect fitting to the estimator rathe

  2. self definitional [Sec. 2.4 and Sec. 4.2, Table 3]
    "VIC-rfg ... architecturally eliminates the impression leakage by removing the reference audio from the synthesis process. ... a minimal 0.05 for VIC-rfg, demonstrating their superior robustness against impression leakage."

    ΔV is defined as RVI-MSE − VI-MSE and is intended to quantify leakage from the reference r. In VIC-rfg, however, r is replaced by random noise, so the synthesized audio s(r,v) does not depend on r at all. The near-zero ΔV is therefore an architectural consequence of removing the reference input; it is not an experimentally demonstrated suppression of leakage. Reporting this value as evidence of robustness is tautological: the quantity the metric was designed to detect was eliminated by construction, and the residual difference reflects only the different target VI values, not reference influence.

full rationale

The corpus contribution is real and externally anchored to LibriTTS-R; the release of manual annotations and estimated VIs is not circular. The methods are also genuine architectural proposals. However, the central quantitative claim of improved controllability rests on the VIE, which is used twice (as training-target generator and as evaluation metric). The paper's own Table 1 documents low inter-annotator agreement, and the subjective evaluation uses the same four annotators who created the training labels and reports only one speaker, so it does not fully break the loop. The reference-free method's low ΔV is expected by design. These are partial, not total, circularities: the subjective MOS test with 30 listeners and the external benchmark against LibriTTS-R provide some independent content. But the headline objective and subjective controllability numbers are not independent of the estimator/annotator loop, so a score of 6 is appropriate.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the validity of the 11 perceptual scales, the reliability of the small human annotation set, the accuracy of the VIE as an evaluation instrument, and the assumption that acoustic similarity propagates impression labels. These are domain assumptions rather than proven facts.

free parameters (2)
  • VIE network parameters = trained on 100 manual labels plus augmented similar utterances
    The Voice Impression Estimator is fitted to the annotation data and is used both to generate training targets and to compute objective evaluation metrics, so its accuracy is a direct input to the central claim.
  • Similarity propagation cutoff = up to 100 acoustically similar utterances per annotated utterance
    The hand-chosen number of neighboring utterances assigned the same manual VI label controls the size and quality of the VIE training set; no validation is provided that these neighbors share the target impression.
assumptions (6)
  • domain assumption The adopted 11 VI scales from the original VIC study are valid, human-meaningful axes of voice impression.
    The paper adopts these scales for comparability and does not independently validate the axis set for LibriTTS-R.
  • domain assumption The VIE is a sufficiently accurate proxy for human VI judgments to serve as both training target and evaluation metric.
    Objective results in Eqs. 6 to 8 and Table 3 rely on this proxy; low inter-annotator agreement puts pressure on it.
  • domain assumption Low inter-annotator agreement does not invalidate the labels.
    The paper argues alpha 0.464 is comparable to prior emotion and singing studies, but it is below the conventional reliability threshold, and the labels still drive training.
  • ad hoc to paper Acoustically similar utterances from the same speaker share the manual VI label.
    Sec. 3 propagates a manual label to up to 100 similar utterances using pitch, energy, and WavLM similarity, with no perceptual validation of the propagation.
  • domain assumption HuBERT and style token embeddings can carry speaker identity separately from the target impression.
    VIC-sep and VIC-rfg assume that speaker and VI can be cleanly separated in the conditioning path; if not, identity or controllability degrades.
  • domain assumption The four professional annotators' ratings generalize to a broader population.
    Subjective controllability results use these same four annotators, not a fresh independent listener panel.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LibriTTS-VI: A Public Corpus and Novel Methods for Efficient Voice Impression Control." pith.science (2026). https://pith.science/paper/LPOIODXA

@misc{pith2026250915626,
  author       = {Pith},
  title        = {Pith review of: LibriTTS-VI: A Public Corpus and Novel Methods for Efficient Voice Impression Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LPOIODXA}},
  note         = {Machine review of arXiv:2509.15626}
}
read the original abstract

Numerical voice impression (VI) control (e.g., scaling brightness) enables fine-grained control in text-to-speech (TTS). However, it faces two challenges: no public corpus and impression leakage, where reference audio biases synthesized voice away from the target VI. To address the first challenge, we introduce LibriTTS-VI, the first public VI corpus built on LibriTTS-R. For the second, we hypothesize a single reference causes leakage by entangling speaker identity and VI. To mitigate this, we propose 1) disentangled training with two utterances from the same speaker for speaker and VI conditioning, and 2) a reference-free method controlling the impression solely via target VI. Experimentally, our best method improves controllability: 11-dimensional VI mean squared error drops from 0.61 to 0.41 objectively and 1.15 to 0.92 subjectively. A comparison with a prompt-based TTS reveals imprecise numerical control and entanglement between VI and text semantics, which our methods overcome.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 3 linked inside Pith

  1. [1]

    The pursuit of this controllability ha s evolved through several paradigms

    INTRODUCTION Modern text-to-speech (TTS) synthesis has achieved near-h uman parity in quality [1, 2, 3], shifting the research focus from simple naturalness toward enhancing controllability over speake r identity and speaking styles. The pursuit of this controllability ha s evolved through several paradigms. Control paradigms include dire ct ma- nipulatio...

  2. [2]

    Good morning

    VIC SYSTEMS This section details the systems used in our study. We first de scribe the VITS-based backbone TTS model [3], which serves as the ba se model for all systems in this paper. We then introduce our bas eline system, VIC-base, which is a reimplementation of the origin al VIC framework [23]. Finally, we present our two novel methods, V IC- sep and V...

  3. [3]

    LIBRITTS-VI: A PUBLIC DATASET FOR VIC To foster reproducible research, we created LibriTTS-VI, a new public dataset, built by annotating the LibriTTS-R corpus [ 24]. We follow [23] and define 11 perceptual scales: A) Low-High, B) Masculine-Feminine, C) Clear-Hoarse, D) Calm-Restless , E) Powerful-Weak, F) Y outhful-Aged, G) Thick-Thin, H) Firm-Relaxed, I) ...

  4. [4]

    EXPERIMENTS We evaluated our systems both objectively and subjectively, focusing on controllability and synthesis quality on the LibriTTS-R test-clean, which is composed of 39 unseen speakers during training, to a ssess performance in a zero-shot setting. 4.1. Experimental setup Our implementation followed the original VIC study [23], bu t some hyperparam...

  5. [5]

    CONCLUSION In this paper, we addressed two key challenges in VIC: the iss ue of impression leakage and the lack of a public corpus. To miti - gate impression leakage, we proposed two novel methods: a tr ain- ing strategy using separate utterances for disentanglemen t, and a reference-free generation model that structurally eliminates the leak- age. Our ex...

  6. [6]

    Emotional end-to-end neural speech synth e- sizer,

    Y . Lee et al., “Emotional end-to-end neural speech synth e- sizer,” arXiv preprint arXiv:1711.05447, 2017

  7. [7]

    Natural TTS synthesis by conditioning wa venet on mel spectrogram predictions,

    J. Shen et al., “Natural TTS synthesis by conditioning wa venet on mel spectrogram predictions,” in Proc. of ICASSP , 2018, pp. 4779–4783

  8. [8]

    FastSpeech 2: Fast and high-quality end-t o-end text to speech,

    Y . Ren et al., “FastSpeech 2: Fast and high-quality end-t o-end text to speech,” in Proc. of ICLR, 2021

Show all 45 references
  1. [9]

    Conditional variational autoencoder wit h ad- versarial learning for end-to-end text-to-speech,

    J. Kim et al., “Conditional variational autoencoder wit h ad- versarial learning for end-to-end text-to-speech,” in Proc. of ICML, 2021, pp. 5530–5540

  2. [10]

    Drawspeech: Expressive speech synthesi s using prosodic sketches as control conditions,

    W. Chen et al., “Drawspeech: Expressive speech synthesi s using prosodic sketches as control conditions,” in Proc. of ICASSP, 2025, pp. 1–5

  3. [11]

    VITS2: Improving quality and efficiency o f single-stage text-to-speech with adversarial learning and archi- tecture design,

    J. Kong et al., “VITS2: Improving quality and efficiency o f single-stage text-to-speech with adversarial learning and archi- tecture design,” in Proc. of Interspeech, 2023, pp. 4374–4378

  4. [12]

    Emosphere++: Emotion-controllable zero- shot text-to-speech via emotion-adaptive spherical vecto r,

    D.-H. Cho et al., “Emosphere++: Emotion-controllable zero- shot text-to-speech via emotion-adaptive spherical vecto r,” IEEE Transactions on Affective Computing, 2025

  5. [13]

    EmoSpeech: guiding fastspee ch2 towards emotional text to speech,

    D. Diatlova and V . Shutov, “EmoSpeech: guiding fastspee ch2 towards emotional text to speech,” in Proc. of SSW, 2023, pp. 106–112

  6. [14]

    Transfer learning from speaker verificati on to multispeaker text-to-speech synthesis,

    Y . Jia et al., “Transfer learning from speaker verificati on to multispeaker text-to-speech synthesis,” Advances in NeurIPS , 2018

  7. [15]

    YourTTS: Towards zero-shot multi- speaker TTS and zero-shot voice conversion for everyone,

    E. Casanova et al., “YourTTS: Towards zero-shot multi- speaker TTS and zero-shot voice conversion for everyone,” i n Proc. of ICML, 2022, pp. 2709–2720

  8. [16]

    Style tokens: Unsupervised style model ing, control and transfer in end-to-end speech synthesis,

    Y . Wang et al., “Style tokens: Unsupervised style model ing, control and transfer in end-to-end speech synthesis,” in Proc. of ICML, 2018, pp. 5180–5189

  9. [17]

    EmoSphere-TTS: Emotional style and i n- tensity modeling via spherical emotion vector for controll able emotional text-to-speech,

    D.-H. Cho et al., “EmoSphere-TTS: Emotional style and i n- tensity modeling via spherical emotion vector for controll able emotional text-to-speech,” in Proc. of Interspeech , 2024, pp. 1810–1814

  10. [18]

    V oiceMe: Personalized voice genera tion in TTS,

    P . van Rijn et al., “V oiceMe: Personalized voice genera tion in TTS,” in Proc. of Interspeech, 2022, pp. 2588–2592

  11. [19]

    Prompttts: Controllable text-to-speec h with text descriptions,

    Z. Guo et al., “Prompttts: Controllable text-to-speec h with text descriptions,” in Proc. of ICASSP, 2023, pp. 1–5

  12. [20]

    Promptstyle: Controllable style transf er for text- to-speech with natural language descriptions,

    G. Liu et al., “Promptstyle: Controllable style transf er for text- to-speech with natural language descriptions,” in Proc. of In- terspeech, 2023, pp. 4888–4892

  13. [21]

    Instructtts: Modelling expressive tts in discrete latent space with natural language style prompt,

    D. Yang et al., “Instructtts: Modelling expressive tts in discrete latent space with natural language style prompt,” in IEEE/ACM Trans. Audio, Speech, Lang. Process., 2024, pp. 2913–2925

  14. [22]

    Prompttts++: Controlling speaker i dentity in prompt-based text-to-speech using natural language descrip- tions,

    R. Shimizu et al., “Prompttts++: Controlling speaker i dentity in prompt-based text-to-speech using natural language descrip- tions,” in Proc. of ICASSP, 2024, pp. 12672–12676

  15. [23]

    LibriTTS-P: A corpus with speaking style and speaker identity prompts for text-to-speech and s tyle captioning,

    M. Kawamura et al., “LibriTTS-P: A corpus with speaking style and speaker identity prompts for text-to-speech and s tyle captioning,” in In Proc. of Interspeech, 2024

  16. [24]

    LibriTTS-R: A restored multi-speak er text- to-speech corpus,

    Y . Koizumi et al., “LibriTTS-R: A restored multi-speak er text- to-speech corpus,” in Proc. of Interspeech , 2023, pp. 5496– 5500

  17. [25]

    Controllable generation of artificial sp eaker em- beddings through discovery of principal directions,

    F. Lux et al., “Controllable generation of artificial sp eaker em- beddings through discovery of principal directions,” in Proc. of Interspeech, 2023, pp. 4788–4792

  18. [26]

    Speech synthesis along perceptu al voice quality dimensions,

    F. Rautenberg at al., “Speech synthesis along perceptu al voice quality dimensions,” in Proc. of ICASSP, 2025, pp. 1–5

  19. [27]

    Extraction of everyday expression asso ciated with voice quality of normal utterance,

    H. Kido et al., “Extraction of everyday expression asso ciated with voice quality of normal utterance,” Journal of the Acous- tical Society of Japan , pp. 405–411, 1999, (In Japanese with English title)

  20. [28]

    The influence of semantic primitives i n an emotion-mediated willingness to buy model from advertisin g speech,

    M. Nagano et al., “The influence of semantic primitives i n an emotion-mediated willingness to buy model from advertisin g speech,” Acoust. Sci. & Tech., vol. 46, no. 1, pp. 87–95, 2025

  21. [29]

    V oice impression control in zero-sho t TTS,

    K. Fujita et al., “V oice impression control in zero-sho t TTS,” in Proc. of Interspeech, 2025

  22. [30]

    Perceptual (but not acoustic) featur es predict singing voice preferences,

    C. Bruder et al., “Perceptual (but not acoustic) featur es predict singing voice preferences,” Sci. Rep., vol. 14, no. 1, pp. 8977, 2024

  23. [31]

    Hubert: Self-supervised speech repre - sentation learning by masked prediction of hidden units,

    W.N. Hsu et al., “Hubert: Self-supervised speech repre - sentation learning by masked prediction of hidden units,” IEEE/ACM Trans. Audio, Speech, Lang. Process. , vol. 29, pp. 3451–3460, 2021

  24. [32]

    Unsupervised domain adaptation by backpropagation,

    Y . Ganin and V . Lempitsky, “Unsupervised domain adaptation by backpropagation,” in Proc. of ICML, 2015, pp. 1180–1189

  25. [33]

    whisper-timestamped,

    J. Louradour., “whisper-timestamped,” https://github.com/linto-ai/whisper-timestamped , 2023

  26. [34]

    Krippendorff, Content Analysis: An Introduction to Its Methodology, SAGE Publications, 2018

    K. Krippendorff, Content Analysis: An Introduction to Its Methodology, SAGE Publications, 2018

  27. [35]

    Ramas: Russian multimodal corp us of dyadic interaction for affective computing,

    O. Perepelkina et al., “Ramas: Russian multimodal corp us of dyadic interaction for affective computing,” in Proc. of SPECOM, 2018, pp. 501–510

  28. [36]

    Scalable diffusion models with tran sform- ers,

    W. Peebles et al., “Scalable diffusion models with tran sform- ers,” in Proc. of ICCV, 2023, pp. 4195–4205

  29. [37]

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

    S. Chen et al., “WavLM: Large-scale self-supervised pr e- training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing, pp. 1505–1518, 2022

  30. [38]

    Conformer: Convolution-augmented t rans- former for speech recognition,

    A. Gulati et al., “Conformer: Convolution-augmented t rans- former for speech recognition,” in Proc. of Interspeech, 2020, pp. 5036–5040

  31. [39]

    ESPnet2-TTS: Extending the edge of T TS research,

    T. Hayashi et al., “ESPnet2-TTS: Extending the edge of T TS research,” arXiv:2110.07840, 2021

  32. [40]

    HierSpeech: Bridging the gap between text and speech by hierarchical variational inference using sel f- supervised representations for speech synthesis,

    S.-H. Lee et al., “HierSpeech: Bridging the gap between text and speech by hierarchical variational inference using sel f- supervised representations for speech synthesis,” in Proc. of NeurIPS, 2022, pp. 16624–16636

  33. [41]

    HierSpeech++: Bridging the gap betwe en semantic and acoustic representation of speech by hierarch i- cal variational inference for zero-shot speech synthesis,

    S.-H. Lee et al., “HierSpeech++: Bridging the gap betwe en semantic and acoustic representation of speech by hierarch i- cal variational inference for zero-shot speech synthesis, ” IEEE Trans. Neural Netw. Learn. Syst., pp. 1–15, 2025

  34. [43]

    Robust speech recognition via large -scale weak supervision,

    A. Radford et al., “Robust speech recognition via large -scale weak supervision,” in Proc. of ICML. PMLR, 2023, pp. 28492– 28518

  35. [44]

    UTMOS: Utokyo-sarulab system for voic emos challenge 2022,

    T. Saeki et al., “UTMOS: Utokyo-sarulab system for voic emos challenge 2022,” in Proc. of Interspeech, 2022, pp. 4521–4525

  36. [45]

    Resemblyzer,

    G. Louppe, “Resemblyzer,” https://github.com/resemble-ai/Resemblyzer, 2019

  37. [8555]

    VI Condition base sep rfg A Mod

    (Bold/dagger (†): significantly better/worse than ba seline, p < 0.05). VI Condition base sep rfg A Mod. -3 3.31 ± 0.10 3.42 ± 0.11 3.75 ± 0.11 Mod. +0 3.31 ± 0.12 3.46 ± 0.11 3.44 ± 0.12 Mod. +3 3.44 ± 0.12 3.21 ± 0.11† 2.85 ± 0.11† B Mod. -3 3.64 ± 0.11 3.42 ± 0.13† 3.46 ± 0....

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.