Pith. sign in

REVIEW 4 major objections 5 minor 44 references

O_O-VC: Synthetic Data-Driven One-to-One Alignment for Any-to-Any Voice Conversion

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Voice conversion can be trained as a direct speech-to-speech mapping on synthetic pairs from a multi-speaker TTS, eliminating disentanglement; this yields zero-shot conversion with 16.35% fewer word errors and 5.91% higher speaker similarit

desk verdict Genuinely novel synthetic-pair training idea, but the missing 'no-synthetic' ablation and the LibriSpeech fine-tuning confound undercut the headline gains. read the letter →

arxiv 2510.09061 v1 pith:QV3M37PT submitted 2025-10-10 cs.SD eess.AS

classification cs.SDeess.AS
keywords syntheticdatavoiceconversionzero-shotmulti-speakerTTSVITSF0conditioningspeakerdisentanglementtwo-phasefine-tuning
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

The paper proposes replacing the usual disentanglement objective in voice conversion with a supervised input-output mapping learned from synthetic data. A pretrained multi-speaker TTS synthesizes pairs of utterances that share the same text, latent linguistic code, duration, and prosody but differ only in speaker identity; these perfectly aligned pairs teach the model to convert voices directly. A two-phase schedule (synthetic paired training, then unlabeled real-speech fine-tuning) plus an F0 conditioning path lets the model handle unseen speakers and new languages. If correct, the method reframes zero-shot voice conversion as a well-posed speech-to-speech regression problem rather than a reconstruction problem.

What carries the argument

The shared-latent pair generator: VITS's text encoder produces h_text, a duration predictor sets a single duration from a randomly chosen speaker token, and a length regulator expands the linguistic posterior to that duration; sampling one z_p and passing it through the inverse flow under two different speaker embeddings yields source and target audios that are aligned frame-by-frame. This is what makes a direct conversion mapping learnable. A second piece is the F0 encoder, which feeds the target pitch contour into the decoder to bridge the pitch difference between the two speakers, and the two-phase training schedule.

What would settle it

Train the same two-phase pipeline but with synthetic pairs generated from different texts (so content differs); if WER and SECS stay similar, frame alignment is not the driver. Alternatively, train a speaker classifier on the shared latent z_p: if it identifies the source speaker above chance, the premise that pairs differ only in speaker identity fails. A third check: fine-tune phase 2 on a corpus whose speakers are disjoint from LibriSpeech test-clean and see whether the SECS gain persists.

Watch

Extended reading notes

Core claim

The central discovery is a training-data construction. Using VITS, the authors sample a single linguistic latent z_p and expand it to frame length; then they invert the flow conditioned on a source speaker and on a target speaker to get two acoustic representations that share content, duration, and prosody, and decode them to audio. Training a FreeVC-style backbone on these as input-output pairs, with target F0 provided to the decoder, teaches a direct mapping that avoids speaker leakage. Experiments on LibriSpeech test-clean and on Chinese, Italian, and Vietnamese show the resulting model beats several published systems in intelligibility and speaker similarity and adapts to new languages w

Load-bearing premise

Everything rests on the public VITS model being able to synthesize two utterances from the same text that share the same linguistic latent, duration, and prosody and differ only in speaker identity; if speaker information leaks into the shared latent or durations diverge, the learned mapping is misaligned and the reported gains would shrink.

Editorial extensions

If this is right

  • Voice conversion training no longer requires parallel human recordings, transcripts, or forced alignment; any multi-speaker TTS with a shared latent can generate unlimited aligned pairs.
  • The content encoder learns to discard speaker identity more thoroughly because the synthetic pairs are identical except for voice, directly reducing speaker leakage.
  • Phase-2 fine-tuning on unlabeled real speech is sufficient to adapt the model to unseen speakers and entirely new languages, suggesting a recipe for low-resource VC.
  • Model performance is upper-bounded by the quality and alignment fidelity of the TTS; replacing VITS with a better multi-speaker TTS should improve conversion.

Reading between the lines

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

  • Editorial: The apparent 5.91% SECS gain may partly reflect phase-2 fine-tuning on LibriSpeech, which overlaps the test set's speaker population; a test using a disjoint fine-tuning corpus would isolate the synthetic-data contribution.
  • Editorial: The method implies a general recipe — any TTS with a controllable speaker-conditioned latent and shared prosody could serve as a data generator; a valuable stress test is whether a weaker TTS with less duration control degrades conversion proportionally.
  • Editorial: The frame-alignment heatmap in the paper suggests the pairs are semantically aligned, but it does not prove z_p is speaker-free; a direct probe (training a speaker classifier on z_p) would test the load-bearing assumption.
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 / 5 minor

Summary. The paper proposes O_O-VC, a zero-shot any-to-any voice conversion method that avoids audio reconstruction and feature disentanglement by training on synthetic parallel pairs generated by a pretrained multi-speaker VITS model. For a given text, the TTS synthesizes a source and a target utterance that are claimed to share linguistic content, duration, and prosody while differing only in speaker identity. The VC model uses a FreeVC-style backbone with a frozen WavLM content extractor, a speaker encoder, and an F0 encoder, and is trained in two phases: first on synthetic VCTK-like pairs, then fine-tuned on LibriSpeech with a reconstruction objective. Evaluation on LibriSpeech test-clean reports the best WER, CER, SECS, and a competitive NISQA, and the paper includes ablation, pitch-correlation, speaker-removal, cross-lingual, emotion-preservation, and semantic-alignment analyses. The central claim is that synthetic-data training plus F0 conditioning plus two-phase fine-tuning yields gains of 16.35% relative WER reduction and 5.91% SECS improvement over several off-the-shelf baselines.

Significance. The central idea—using synthetic same-content/different-speaker pairs from a pretrained TTS as direct input-output supervision—is a clean and potentially valuable alternative to disentanglement-based VC. If the reported gains were cleanly attributable to this training signal, the paper would be a solid contribution: it avoids label noise, removes the need for forced alignment, and offers a scalable way to create paired VC data. The paper also provides a reasonable set of objective and subjective measures, an explicit zero-shot protocol, and multilingual adaptation experiments. However, the current experimental design does not yet establish the key attribution: the best configuration is also the only one fine-tuned on the evaluation domain, and the promised control without synthetic data is missing. The manuscript is therefore promising but needs a dedicated control experiment before the central claim can be accepted.

major comments (4)
  1. [§5.2, Table 2] The text states that 'removing the use of synthetic data, the F0 encoder, or phase 2 fine-tuning each leads to a noticeable drop in intelligibility,' but Table 2 contains no 'w/o synthetic data' row. The only ablation rows are 'w/o F0 Encoder' and 'w/o Finetuning.' This missing control is load-bearing because the existing 'w/o Finetuning' row points the other way: removing phase-2 LibriSpeech fine-tuning drops SECS from 86.70 to 70.78, which is worse than FreeVC's 75.66, and raises WER from 1.74 to 2.18. Thus the synthetic-only model is comparable to FreeVC on WER but substantially worse on speaker similarity. Without a model trained on the same real corpus in phase 1 while omitting synthetic data, the paper cannot support the claim that synthetic pairs are what produce the headline gains.
  2. [§4.1–4.2, §5.1, Table 1] The comparison in Table 1 is confounded by domain adaptation. O_O-VC is fine-tuned on LibriSpeech train-clean-100 and train-clean-360 and evaluated on LibriSpeech test-clean, while all baselines are used with their official off-the-shelf pretrained weights. Therefore the reported 16.35% relative WER reduction and 5.91% SECS improvement over FreeVC may reflect in-domain fine-tuning rather than the proposed synthetic-pair phase. A fair control is required: either fine-tune the same backbone on LibriSpeech without phase-1 synthetic data, or also fine-tune each baseline under the same domain-matched protocol. Without this, the central attribution to synthetic data is unsupported.
  3. [§3.1, Eqs. (1)–(9)] The entire training signal depends on the ability of the official pretrained VITS to generate source and target utterances from the same linguistic latent z_p, the same duration, and the same prosody while differing only in speaker identity. The paper does not demonstrate that the publicly released VITS model can actually be operated this way: the equations require sharing dur_text and the sampled z_p across the two speaker-conditioned inverse-flow paths, but standard VITS inference predicts duration per speaker and does not expose a straightforward mechanism for forcing identical duration/prosody across arbitrary speaker pairs. If the official implementation cannot enforce this, the synthetic pairs are not one-to-one aligned and the learned mapping inherits that misalignment. The paper should provide either implementation-level details of how Eqs. (6)–(9) were realized with the public c
  4. [§3.1, §3.2.1] The paper asserts that synthetic pairs differ 'only in speaker identity' while also stating in §3.2.1 that speaker identity 'primarily manifests in variations in pitch' and that an F0 encoder is needed to handle the F0 mismatch between source and target. These statements are in tension: if the synthetic pairs are truly identical in prosody, their F0 contours should also match. The role of the F0 encoder and the precise claim about what is shared (segmental duration? intonation? F0?) should be clarified. This ambiguity affects the interpretation of the w/o F0 Encoder ablation and the F0-PCC analysis.
minor comments (5)
  1. [§5.3, Table 4] The table header reads 'CER/WER' but the rows do not indicate which metric applies to which language. The text mentions CER for Chinese and WER for Vietnamese and Italian; make this explicit in the table or caption.
  2. [§5.2, Figure 3] The F0-PCC figure is mentioned in the text but no quantitative values are reported in the table; reporting the exact numbers would make the pitch-preservation claim easier to verify.
  3. [§4.3] The paper introduces B-MOS as a new summary metric. Its motivation is clear enough, but it would help to state explicitly that it is the arithmetic mean of MOS and SMOS, and to note that it is not an independently validated perceptual score.
  4. [Appendix A.2] The phrase 'native-like qualities' is stronger than the measured quantity: CER/WER between converted audio and ground-truth transcriptions is an intelligibility measure, not a nativeness measure. Please reword.
  5. [General] No code or trained model is released, only a demo page. For a method whose central mechanism depends on a nonstandard use of a public VITS checkpoint, releasing the data-generation script would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the synthetic-pair training pipeline is an empirical method with a missing-control caveat, not a by-construction reduction.

full rationale

The paper's derivation chain does not reduce to its inputs. Synthetic pairs are generated by a pretrained multi-speaker VITS from a shared text and shared latent content with different speaker conditions (Eqs. 1-9); these pairs are then used as external supervision to train a separate VC model with reconstruction/conversion, KL, GAN, and distillation losses (Eqs. 14-15). No reported metric (WER, CER, SECS) is a fitted parameter, and the semantic-alignment heatmap in Sec. 5.4 is a diagnostic sanity check of the data-generation scheme, not a predicted quantity. The only self-citation, Tu et al. (2025), is used for Vietnamese dataset provenance in Sec. 5.3 and Appendix A.2, and it is not load-bearing for the central claim. The main weakness is a support gap rather than circularity: Sec. 5.2 claims that removing synthetic data leads to a noticeable drop, but Table 2 contains no 'w/o synthetic data' row; the closest ablation, 'w/o Finetuning,' shows a large SECS drop (86.70 to 70.78) and only a small WER change (1.74 to 2.18). This leaves the contribution of phase-1 synthetic data versus phase-2 LibriSpeech fine-tuning under-identified, but that is an omitted-control/attribution problem, not an equivalence by construction.

Assumptions & free parameters 0 free parameters · 5 assumptions · 1 invented entities

The central claim depends on the validity of synthetic-pair alignment from VITS, on SSL content extraction, and on the fairness of the LibriSpeech evaluation. No closed-form free parameters are fitted; the method's learned weights are standard ML model parameters.

assumptions (5)
  • domain assumption A pretrained multi-speaker VITS model can generate source and target utterances with identical linguistic latent, duration, and prosody from a shared z_p (Eq. 5-9).
    Central to the synthetic-pair training signal; no independent verification beyond the paper's own Figure 5.
  • domain assumption WavLM plus a content extractor yields a speaker-independent content distribution from real source audio.
    Section 3.2.1; standard SSL assumption, but speaker leakage is exactly the issue the paper claims to solve, so this premise is load-bearing and unproven for real speech.
  • domain assumption Phase-2 reconstruction fine-tuning on LibriSpeech preserves the speaker-independent features learned in phase 1.
    Section 3.2.2; freezing WavLM/content extractor is asserted to protect learned representations, with no controlled experiment isolating this.
  • domain assumption LibriSpeech test-clean speakers are truly unseen and the zero-shot evaluation is representative.
    Section 4.1/4.3; train/test speaker disjointness is standard for LibriSpeech but not demonstrated in the paper, and the model was fine-tuned on LibriSpeech train while baselines were not.
  • domain assumption HuBERT-based WER/CER and Resemblyzer SECS are valid proxies for intelligibility and speaker similarity.
    Section 4.3; these are standard tools, but objective scores are reported without confidence intervals.
invented entities (1)
  • Balance-MOS (B-MOS)
    purpose: Composite subjective metric averaging MOS and SMOS into a single score for system ranking.
    Defined in Section 4.3 as the mean of two subjective scores; no external validation, and it is not used in the headline WER/SECS claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of O_O-VC: Synthetic Data-Driven One-to-One Alignment for Any-to-Any Voice Conversion." pith.science (2026). https://pith.science/paper/QV3M37PT

@misc{pith2026251009061,
  author       = {Pith},
  title        = {Pith review of: O_O-VC: Synthetic Data-Driven One-to-One Alignment for Any-to-Any Voice Conversion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QV3M37PT}},
  note         = {Machine review of arXiv:2510.09061}
}
read the original abstract

Traditional voice conversion (VC) methods typically attempt to separate speaker identity and linguistic information into distinct representations, which are then combined to reconstruct the audio. However, effectively disentangling these factors remains challenging, often leading to information loss during training. In this paper, we propose a new approach that leverages synthetic speech data generated by a high-quality, pretrained multispeaker text-to-speech (TTS) model. Specifically, synthetic data pairs that share the same linguistic content but differ in speaker identity are used as input-output pairs to train the voice conversion model. This enables the model to learn a direct mapping between source and target voices, effectively capturing speaker-specific characteristics while preserving linguistic content. Additionally, we introduce a flexible training strategy for any-to-any voice conversion that generalizes well to unseen speakers and new languages, enhancing adaptability and performance in zero-shot scenarios. Our experiments show that our proposed method achieves a 16.35% relative reduction in word error rate and a 5.91% improvement in speaker cosine similarity, outperforming several state-of-the-art methods. Voice conversion samples can be accessed at: https://oovc-emnlp-2025.github.io/

Figures

Figures reproduced from arXiv: 2510.09061 by the authors.

Figure 1
Figure 1. Voice conversion with synthetic data. 3.2.1 Training Procedure In the training phase, the source and target audios are processed through different stages: Source Audio Processing: The source audio is passed through a pretrained WavLM (Chen et al., 2022) and a content extractor to obtain a distribu￾tion of content features N (µθ, σ2 θ ). Target Audio Processing: The target audio is passed through a speaker encoder an… view at source ↗
Figure 2
Figure 2. T-SNE visualization of speaker-independent features. More distributed points with no clusters indicate [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Comparison of systems on F0-PCC clustering evaluation metrics: Adjusted Rand In￾dex (ARI), Normalized Mutual Information (NMI) and Silhouette Score. The ARI measures the simi￾larity between predicted clusters and true speaker labels, adjusted for chance. A lower ARI indicates that the clusters do not correspond well to speaker identities, suggesting better speaker information removal. NMI measures the amount of shar… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance of new language adaptation: CER for Chinese, WER for Vietnamese and Italian. 5.4 Semantic Alignment of Synthetic Audio Pairs (a) Cosine pairwise semantic similarity. (b) Top-1 cosine similarity alignment path [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Semantic alignment of source and target audio [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Emotion representation of converted audio [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 5 canonical work pages

  1. [1]

    Matthew Baas, Benjamin van Niekerk, and Herman Kamper. 2023. Voice conversion with just nearest neighbors. In Interspeech 2023

  2. [2]

    Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. 2009. https://doi.org/10.1007/978-3-642-00296-0_5 Pearson correlation coefficient . In Noise Reduction in Speech Processing, pages 1--4, Berlin, Heidelberg. Springer Berlin Heidelberg

  3. [3]

    Ling-Hui Chen, Zhen-Hua Ling, Li-Juan Liu, and Lirong Dai. 2014. https://doi.org/10.1109/TASLP.2014.2353991 Voice conversion using deep neural networks with layer-wise generative training . Audio, Speech, and Language Processing, IEEE/ACM Transactions on, 22:1859--1872

  4. [4]

    Nanxin Chen, Yu Zhang, Heiga Zen (Byungha Chun), Ron Weiss, Mohammad Norouzi, and William Chan. 2021 a . https://arxiv.org/abs/2009.00713 Wavegrad: Estimating gradients for waveform generation . In ICLR

  5. [5]

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Michael Zeng, Xiangzhan Yu, and Furu Wei. 2022. https://doi.org/10.1109/JSTSP.2022.3188113 Wavlm: Large-scale self-supervised pre-training for full stack speech processing . IE...

  6. [7]

    Ha-Yeong Choi, Sang-Hoon Lee, and Seong-Whan Lee. 2023. https://doi.org/10.21437/Interspeech.2023-817 Diff-hiervc: Diffusion-based hierarchical voice conversion with robust pitch generation and masked prior for zero-shot speaker adaptation . In Interspeech 2023, pages 2283--2287

  7. [8]

    Ha-Yeong Choi, Sang-Hoon Lee, and Seong-Whan Lee. 2024. Dddm-vc: Decoupled denoising diffusion models with disentangled representation and prior mixup for verified robust voice conversion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17862--17870

  8. [9]

    Zhifu Gao, Zerui Li, Jiaming Wang, Haoneng Luo, Xian Shi, Mengzhe Chen, Yabin Li, Lingyun Zuo, Zhihao Du, Zhangyu Xiao, and Shiliang Zhang. 2023. Funasr: A fundamental end-to-end speech recognition toolkit. In INTERSPEECH

Show all 44 references
  1. [10]

    Zhifu Gao, Shiliang Zhang, Ian Mcloughlin, and Zhijie Yan. 2022. Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to-end speech recognition. In Interspeech 2022, pages 2063--2067

  2. [11]

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. https://doi.org/10.1109/TASLP.2021.3122291 Hubert: Self-supervised speech representation learning by masked prediction of hidden units . IEEE/ACM Transactions on ...

  3. [12]

    Hao Huang, Lin Wang, Jichen Yang, Ying Hu, and Liang He. 2023. W2vc: Wavlm representation based one-shot voice conversion with gradient reversal distillation and ctc supervision. In Journal AUDIO SPEECH MUSIC PROC

  4. [13]

    Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. 2022 a . Fastdiff: A fast conditional diffusion model for high-quality speech synthesis. In Proceedings of the IJCAI-22

  5. [14]

    Wen-Chin Huang, Shu-Wen Yang, Tomoki Hayashi, Hung-Yi Lee, Shinji Watanabe, and Tomoki Toda. 2022 b . S3prl-vc: Open-source voice conversion framework with self-supervised speech representations. In Proc. ICASSP

  6. [15]

    Wen-Chin Huang, Shu-Wen Yang, Tomoki Hayashi, and Tomoki Toda. 2022 c . A Comparative Study of Self-Supervised Speech Representation Based Voice Conversion . IEEE Journal of Selected Topics in Signal Processing, 16(6):1308--1318

  7. [17]

    Zeqian Ju, Yuancheng Wang, Kai Shen, Xu Tan, Detai Xin, Dongchao Yang, Yanqing Liu, Yichong Leng, Kaitao Song, Siliang Tang, Zhizheng Wu, Tao Qin, Xiang-Yang Li, Wei Ye, Shikun Zhang, Jiang Bian, Lei He, Jinyu Li, and Sheng Zhao. 2024. Naturalspeech 3: zero-shot speech synthes...

  8. [18]

    Joonson Kim, Jungil Kong, and Jaehyeon Son. 2021. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In Proceedings of the International Conference on Machine Learning (ICML), pages 5530--5540

  9. [19]

    Ziyu Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. 2021. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations (ICLR)

  10. [20]

    Larochelle, and Ole Winther

    Anders Boesen Lindbo Larsen, S ren Kaae S nderby, H. Larochelle, and Ole Winther. 2016. Autoencoding beyond pixels using a learned similarity metric. International conference on machine learning, abs/1512.09300

  11. [22]

    Jingyi Li, Weiping Tu, and Li Xiao. 2023. https://doi.org/10.1109/ICASSP49357.2023.10095191 Freevc: Towards high-quality text-free one-shot voice conversion . In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1--5

  12. [23]

    Lin, Chung-Ming Chien, Jheng-Hao Lin, Hung-yi Lee, and Lin-shan Lee

    Yist Y. Lin, Chung-Ming Chien, Jheng-Hao Lin, Hung-yi Lee, and Lin-shan Lee. 2021. https://doi.org/10.1109/ICASSP39728.2021.9413699 Fragmentvc: Any-to-any voice conversion by end-to-end extracting and fusing fine-grained voice fragments with attention . In ICASSP 2021 - 2021 I...

  13. [24]

    Songxiang Liu, Yuewen Cao, Disong Wang, Xixin Wu, Xunying Liu, and Helen Meng. 2021. https://doi.org/10.1109/TASLP.2021.3076867 Any-to-many voice conversion with location-relative sequence-to-sequence modeling . IEEE/ACM Transactions on Audio, Speech, and Language Processing

  14. [25]

    Hieu-Thi Luong and Hai-Quan Vu. 2016. https://aclanthology.org/W16-5207 A non-expert K aldi recipe for V ietnamese speech recognition system . In WLSI / OIAF 4 HLT 2016 , pages 51--55, Osaka, Japan. The COLING 2016 Organizing Committee

  15. [26]

    Ziyang Ma, Zhisheng Zheng, Jiaxin Ye, Jinchao Li, Zhifu Gao, Shiliang Zhang, and Xie Chen. 2024. emotion2vec: Self-supervised pre-training for speech emotion representation. Proc. ACL 2024 Findings

  16. [27]

    X. Mao, Q. Li, H. Xie, R. K. Lau, Z. Wang, and S. Smolley. 2017. https://doi.org/10.1109/ICCV.2017.304 Least squares generative adversarial networks . In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2813--2821, Los Alamitos, CA, USA. IEEE Computer Society

  17. [28]

    Gabriel Mittag, Babak Naderi, Assmaa Chehadi, and Sebastian Möller. 2021. https://doi.org/10.21437/interspeech.2021-299 Nisqa: A deep cnn-self-attention model for multidimensional speech quality prediction with crowdsourced datasets . Interspeech 2021

  18. [29]

    Paarth Neekhara, Shehzeen Hussain, Rafael Valle, Boris Ginsburg, Rishabh Ranjan, Shlomo Dubnov, Farinaz Koushanfar, and Julian McAuley. 2024. Selfvc: voice conversion with iterative refinement using self transformations. In Proceedings of the 41st International Conference on M...

  19. [30]

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. https://doi.org/10.1109/ICASSP.2015.7178964 Librispeech: An asr corpus based on public domain audio books . In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), page...

  20. [31]

    Seung-won Park, Doo-young Kim, and Myun-chul Joe. 2020. https://doi.org/10.21437/Interspeech.2020-1542 Cotatron: Transcription-guided speech encoder for any-to-many voice conversion without parallel data . In Proc. Interspeech 2020, pages 4696--4700

  21. [32]

    Adam Polyak, Yossi Adi, Jade Copet, Eugene Kharitonov, Kushal Lakhotia, Wei-Ning Hsu, Abdelrahman Mohamed, and Emmanuel Dupoux. 2021. Speech Resynthesis from Discrete Disentangled Self-Supervised Representations . In Proc. Interspeech 2021

  22. [33]

    Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. 2021. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning

  23. [34]

    Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, Mikhail Sergeevich Kudinov, and Jiansheng Wei. 2022. https://openreview.net/forum?id=8c50f-DoWAu Diffusion-based voice conversion with fast maximum likelihood sampling scheme . In International Conference on Learning...

  24. [35]

    Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. Mls: A large-scale multilingual dataset for speech research. ArXiv, abs/2012.03411

  25. [36]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning, ICML'23

  26. [37]

    Siyuan Shan, Yang Li, Amartya Banerjee, and Junier Oliva. 2024. https://doi.org/10.1609/aaai.v38i13.29411 Phoneme hallucinator: One-shot voice conversion via set expansion . Proceedings of the AAAI Conference on Artificial Intelligence, 38:14910--14918

  27. [38]

    Yao Shi, Hui Bu, Xin Xu, Shaoji Zhang, and Ming Li. 2021. https://doi.org/10.21437/Interspeech.2021-755 Aishell-3: A multi-speaker mandarin tts corpus . In Proceedings of Interspeech, pages 2756--2760

  28. [39]

    Lifa Sun, Shiyin Kang, Kun Li, and Helen Meng. 2015. https://doi.org/10.1109/ICASSP.2015.7178896 Voice conversion using deep bidirectional long short-term memory based recurrent neural networks . In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

  29. [40]

    Lifa Sun, Kun Li, Hao Wang, Shiyin Kang, and Helen Meng. 2016. Phonetic posteriorgrams for many-to-one voice conversion without parallel data training. In 2016 IEEE International Conference on Multimedia and Expo (ICME), pages 1--6

  30. [41]

    Xiaohai Tian, Junchao Wang, Haihua Xu, Eng Chng, and Haizhou Li. 2018. https://doi.org/10.21437/Odyssey.2018-32 Average modeling approach to voice conversion with non-parallel data . In Odyssey 2018 The Speaker and Language Recognition Workshop, pages 227--232

  31. [42]

    Huu Tuong Tu, Luong Thanh Long, Vu Huan, Nguyen Thi Phuong Thao, Nguyen Van Thang, Nguyen Tien Cuong, and Nguyen Thi Thu Trang. 2025. https://doi.org/10.1109/ICASSP49660.2025.10889083 Voice conversion for low-resource languages via knowledge transfer and domain-adversarial tra...

  32. [43]

    Da-Yi Wu and Hung-yi Lee. 2020. https://doi.org/10.1109/ICASSP40776.2020.9053854 One-shot voice conversion by vector quantization . In ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

  33. [44]

    Mingyang Zhang, Yi Zhou, Li Zhao, and Haizhou Li. 2021. https://doi.org/10.1109/TASLP.2021.3066047 Transfer learning from speech synthesis to voice conversion with non-parallel training data . IEEE/ACM Transactions on Audio, Speech, and Language Processing, PP:1--1

  34. [45]

    Kun Zhou, Berrak Sisman, Rui Liu, and Haizhou Li. 2021. Seen and unseen emotional style transfer for voice conversion with a new emotional speech dataset. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 920--924. IEEE

  35. [46]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  36. [47]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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