Zero-Shot Text-to-Speech for Vietnamese

Dat Quoc Nguyen; Linh The Nguyen; Thi Vu

REVIEW 3 major objections 5 minor 31 references

Zero-Shot Text-to-Speech for Vietnamese

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read PhoAudiobook, a 941-hour Vietnamese audiobook corpus with normalized transcripts and speaker labels, improves zero-shot text-to-speech across models, with an XTTS-v2 model trained on it beating the viXTTS baseline on every metric and test…

desk verdict The dataset is real and useful, but the headline comparison (XTTS-v2PAB vs viXTTS) conflates data size, domain, and preprocessing, so the 'data quality drives gains' claim is underdetermined. read the letter →

arxiv 2506.01322 v1 pith:F6PUKEMD submitted 2025-06-02 cs.CL cs.SDeess.AS

Thi Vu , Linh The Nguyen , Dat Quoc Nguyen This is my paper

classification cs.CLcs.SDeess.AS

keywords Vietnamesetext-to-speechzero-shotTTSaudiobookdatasetPhoXTTS-v2VALL-EVoiceCraftspeaker-labeledspeechdata

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 introduces PhoAudiobook, a 941-hour Vietnamese audiobook dataset built by extracting vocal tracks, transcribing them with two independent ASR systems and keeping only exact matches, filtering out multi-speaker audio, normalizing texts, and capping each speaker at four hours. The authors train three zero-shot text-to-speech models (VALL-E, VoiceCraft, and XTTS-v2) on this data plus 554 hours of short-clip augmentation. They report that the XTTS-v2 model trained on PhoAudiobook outperforms viXTTS, the same model fine-tuned on the viVoice dataset, across every objective and subjective metric and every test set, including viVoice's own test split. A reader would care because low-resource languages like Vietnamese lack the large, clean, speaker-labeled datasets that zero-shot TTS needs, and the paper argues PhoAudiobook fills that gap and is publicly released.

What carries the argument

The load-bearing object is the dataset and its creation pipeline: vocal-track separation to remove background audio, transcription by a large multilingual ASR system, keeping only segments where a second Vietnamese ASR system produces an exact transcript match, filtering out multi-speaker clips, normalizing texts with a sequence-to-sequence model, capping each speaker at four hours, and manually verifying the test splits. Long 10–20 second samples are constructed by merging pause-aligned short segments, and the training set is augmented with 554 hours of unmerged short clips so models also see short inputs.

What would settle it

Train XTTS-v2 on the viVoice corpus processed through the same pipeline (volume normalization, speaker IDs, text normalization, and short-clip augmentation) for the same number of steps or epochs, and compare against XTTS-v2PAB; if the gap vanishes, the advantage is not specific to PhoAudiobook. Alternatively, train on PhoAudiobook without the 554-hour augmentation and check whether it still beats viXTTS.

Watch

Extended reading notes

Core claim

The central claim is that a carefully curated long-form audiobook corpus, with transcripts normalized to written Vietnamese and explicit speaker identities, is a better foundation for zero-shot Vietnamese TTS than the existing viVoice corpus, even though viVoice is similar in total hours. The evidence is a full comparison of VALL-E, VoiceCraft, and XTTS-v2 trained on PhoAudiobook against viXTTS, a baseline fine-tuned from the same XTTS-v2 checkpoint on viVoice. On PhoAudiobook's speaker test sets, on VIVOS, and on a held-out viVoice set, XTTS-v2PAB achieves lower WER, MCD, and F0 error, and higher MOS and SMOS. The paper also claims that VALL-E and VoiceCraft handle short sentences better than XTTS-v2 does, attributing XTTS-v2's short-input degradation to an architectural tendency rather than to data.

Load-bearing premise

The paper assumes that the performance gap between XTTS-v2PAB and viXTTS is caused by PhoAudiobook's higher data quality, even though the two models are trained on different amounts of audio (1494 vs 1016 hours) with different segment lengths and preprocessing; a matched-data control is not reported.

Editorial extensions

If this is right

PhoAudiobook is released publicly for non-commercial use, giving Vietnamese TTS research a large speaker-labeled resource with normalized transcripts.
XTTS-v2 fine-tuned on PhoAudiobook beats viXTTS on all reported metrics and test sets, including viVoice's own test audio.
The dataset creation pipeline transfers to other languages that have raw audiobook audio and an ASR model for transcript confirmation.
VALL-E and VoiceCraft trained on PhoAudiobook outperform XTTS-v2 on the short-sentence VIVOS test set, pointing to an XTTS-v2-specific short-input weakness rather than a data issue.
The paper suggests that long-form audiobook data alone is insufficient for short utterances, motivating the short-clip augmentation step.

Reading between the lines

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

The reported advantage of XTTS-v2PAB over viXTTS conflates dataset curation with training duration: XTTS-v2PAB saw 1494 hours (base plus 554 hours of short clips) while viXTTS saw 1016 hours, so a matched-hours ablation is needed to attribute the gain to PhoAudiobook's quality.
If the exact-match cross-ASR filtering is the key step, then the same pipeline could benefit other low-resource languages that have raw audiobook audio but no clean TTS corpus.
A direct test of the short-sentence claim: evaluate VALL-E and VoiceCraft on PhoAudiobook's long samples versus VIVOS's short samples to see whether their advantage is a general short-input property or specific to VIVOS.

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.

Desk Editor's Note

read the letter

Zoe—

The asset here is PhoAudiobook itself: 941 hours of Vietnamese audiobook speech from 735 speakers with explicit IDs, normalized transcripts, 10–20 second segments, and a curation pipeline (demucs, dual-ASR filtering, multi-speaker filtering, silence trimming, volume normalization) described in enough detail to reproduce. For Vietnamese TTS, which has been working with shorter, noisier, often transcription-light corpora, that is a real contribution. The authors also manually inspected the test sets and release the data publicly, which is more than most dataset papers do.

The benchmark section is where I push back. The abstract says PhoAudiobook 'consistently enhances model performance,' and Section 4 says XTTS-v2PAB 'consistently outperforms viXTTS across all metrics and test sets.' The only comparison behind that claim is XTTS-v2PAB trained on 1494 hours of processed audiobook audio versus viXTTS trained on 1016 hours of viVoice with different segment lengths and no transcript normalization. That is several axes of difference, and the paper attributes the whole gap to dataset quality without isolating it. A matched-data control—same audio hours, same preprocessing, only the transcript normalization changed—would make the claim testable. Without it, the gains (4.16 vs 4.23 WER on the seen set) are suggestive at best.

Two smaller issues. WER is computed with PhoWhisper-large, the same ASR used to filter and verify the PhoAudiobook transcripts, so models trained on this data are evaluated by a metric biased toward their own training distribution; an independent ASR would settle it. And Table 2 prints the same WER, 37.81, for XTTS-v2PAB and viXTTS on VIVOS, which contradicts the 'consistently outperforms' sentence as written—likely a typo, but it needs fixing. Objective metrics also come without error bars, which is common but worth noting given the small test sets.

What I like: the short-form finding (VALL-E and VoiceCraft beat XTTS on VIVOS, and both XTTS variants ramble on short inputs) is an architecture-level observation that does not depend on the confounded comparison, and the implementation appendix is unusually complete. The stated limitation—no code-switching evaluation—is honest and proportionate.

Bottom line: the dataset deserves a serious referee and should be cited by anyone working on Vietnamese TTS. The performance attribution needs work first: matched-data control or softened claims, an independent ASR check, and a corrected table. I would send it to review, expecting heavy revision on Section 4.

Referee Report

3 major / 5 minor

Summary. This paper introduces PhoAudiobook, a 941-hour Vietnamese audiobook dataset constructed from 23K hours of raw audiobooks via source separation, ASR transcription, transcript filtering with PhoWhisper-large, multi-speaker filtering, silence trimming, volume normalization, and text normalization. The authors fine-tune VALL-E, VoiceCraft, and XTTS-v2 on the resulting training data (with a short-audio augmentation to 1494 hours) and evaluate against a viXTTS baseline fine-tuned on viVoice, reporting WER, MCD, RMSEF0, MOS, and SMOS on PhoAudiobook seen/unseen speaker test sets, VIVOS, and a viVoice test set. The central claim is that PhoAudiobook consistently enhances performance, with XTTS-v2PAB outperforming viXTTS across all metrics and test sets.

Significance. The dataset release is a potentially valuable contribution: 941 hours of normalized, speaker-labeled Vietnamese audiobook audio with manually inspected test sets, and the pipeline is described in enough detail to be adapted. The experiments span three representative zero-shot TTS architectures and include both seen and unseen speaker conditions. However, the headline comparison is confounded by training-set size, domain, and preprocessing, and the objective evaluation has a circularity risk because the same ASR used for transcript filtering is also used for WER computation. If the confounds were removed, the central claim would be credible; as it stands, the evidence is suggestive but not conclusive.

major comments (3)

[§3.1, §3.2, §4] The central contrast between XTTS-v2PAB and viXTTS does not isolate the contribution of PhoAudiobook's quality. XTTS-v2PAB is trained on 1494 hours (940 h PhoAudiobook plus 554 h of short-audio augmentation reprocessed from the same source), whereas viXTTS is fine-tuned on 1016 hours of viVoice; the two conditions also differ in domain, segment duration distribution, and preprocessing (demucs separation, dual-ASR transcript filtering, volume normalization, text normalization). Any of these factors, most plainly the 47% larger training set, could drive the observed WER/MOS/SMOS differences. The claim in Section 4 that PhoAudiobook 'consistently enhances model performance' is therefore underdetermined. Please add a matched-data control, for example an XTTS-v2 model fine-tuned on a comparable-size subset of PhoAudiobook without the short-audio augmentation, or an XTTS-v2 model fine-tuned on viVoice processed through the same pipeline.
[§3.2, Table 2] The WER metric is computed with PhoWhisper-large, which is the same ASR model used to filter and verify the PhoAudiobook transcripts. PhoAudiobook-trained models are therefore evaluated with a metric that is biased toward their own training transcript distribution, while viXTTS is not. Report WER with at least one independent ASR (e.g., Whisper-large-v3 or a non-PhoWhisper Vietnamese ASR) or justify that the bias is negligible. In addition, objective metrics are reported without error bars or significance tests; the WER difference on PAB-S (4.16 vs. 4.23) is not distinguishable from noise. The duplicate WER value of 37.81 for both XTTS-v2PAB and viXTTS on VIVOS should also be checked, since it contradicts the claim of consistent outperformance.
[§3.2, Table 2, §4] The assertion that XTTS-v2PAB 'consistently outperforms viXTTS across all metrics and test sets' is stronger than the table supports. On VIVOS the two systems have identical WER, and on PAB-S the WER gap is 0.07, which is within the range of what would be expected from test-set sampling without a significance test. The RMSEF0 and MCD values also lack uncertainty estimates, and several MOS/SMOS differences overlap within their reported confidence intervals. Please restrict the claim to the metrics and test sets where the difference is actually large, or provide statistical tests that justify the global statement.

minor comments (5)

[Figure 1 and §2.1] Figure 1 appears to place 'Trim audio & normalize volume' before 'Collect data', while Section 2.1 describes volume normalization after filtering and silence trimming; please align the figure with the textual pipeline order.
[§2.1] The wav2vec2-bartpho model is cited only by a URL; provide a formal reference so that the speaker-filtering step is reproducible.
[§2.2] The SI-SNR comparison is based on 1000 randomly sampled audio files per dataset; state whether the same random seed was used and how the estimator was configured.
[§2.1] The text normalization example uses English ('forty three'); since the dataset is Vietnamese, please give Vietnamese examples (e.g., '43' to 'bốn mươi ba') to clarify the normalization behavior.
[Appendix A] The training duration for VALL-E is not specified; VoiceCraft and XTTS-v2 are given in epochs, but VALL-E only lists GPU count and learning rate. Please report the number of epochs or update steps for all three models.

Circularity Check

1 steps flagged · score 4.0 of 10

WER evidence is partly circular: PhoWhisper-large both curates PhoAudiobook and scores intelligibility, but the central claim still rests on independent metrics.

fitted input called prediction [Section 2.1 (PhoAudiobook creation) and Section 3.2 (Evaluation setup, WER metric)]
"We then retain only the samples where the Whisper-large-v3-based transcription matches exactly with the transcription output from PhoWhisper-large. ... Here, we employ the ASR model PhoWhisper-large (Le et al., 2024) to generate the transcription of the synthesized speech."

PhoWhisper-large is used twice: first as the curation filter that decides which audiobook audio is retained (requiring exact agreement between Whisper-large-v3 and PhoWhisper-large), and second as the ASR that transcribes synthesized speech to compute WER. A model trained on PhoAudiobook is therefore trained on audio-transcript pairs that PhoWhisper already recognizes, and its intelligibility is then scored by that same recognizer. viXTTS, trained on viVoice, was never filtered by PhoWhisper, so the WER gap partly reflects an instrument tuned to PhoAudiobook's selection criterion rather than to intrinsic intelligibility.

full rationale

The paper's central derivation chain is an empirical evaluation, not a formal derivation, so most of the comparison is not circular by construction. The only specific circular component I can exhibit is the WER evaluation loop: PhoWhisper-large is both a data-selection filter in Section 2.1 and the transcription engine for the WER metric in Section 3.2. This creates a real bias in favor of PhoAudiobook-trained models on the intelligibility metric, because the training data were selected to be exactly the samples that PhoWhisper transcribes correctly. However, the paper's claims are not forced by this loop: MCD, RMSEF0, MOS, and SMOS do not use PhoWhisper, and the central XTTS-v2PAB versus viXTTS comparison improves on those metrics too, so there is independent empirical content. I do not count the data-size/domain/preprocessing confound between XTTS-v2PAB (1494 hours, audiobooks, 10-20 s segments) and viXTTS (1016 hours, general-purpose, short segments) as circularity, because that is an experimental control issue, not a definitional reduction. The self-citation of PhoWhisper is also not by itself disqualifying: PhoWhisper is a public, externally available model; the circularity comes from its dual role as curator and evaluator. Overall, a score of 4 reflects a partial circular component in one metric while the central claim retains independent support from the other metrics and external VIVOS/viVoice test sets.

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

PhoAudiobook's utility depends on the reliability of automated curation tools such as speech separation, ASR, and speaker filtering, whose outputs are only manually verified for the small test sets. The experimental claim additionally assumes that a model trained on more audiobook data and evaluated with the same ASR used in curation provides an unbiased comparison with a baseline trained on different data. No formal derivation is involved.

free parameters (5)

Merged segment duration target = 10 to 20 seconds
Hand-chosen in Section 2.1 to provide long-form samples while retaining natural pauses; affects model training distribution.
Minimum transcript length for filtering = 25 words
Hand-chosen cutoff in Section 2.1 to exclude samples with excessive silence; changes dataset composition.
Per-speaker audio cap = 4 hours
Chosen in Section 2.1 for balanced speaker representation; reduces the corpus from 1400 to 941 hours.
Test set sizes = 0.8h seen, 0.5h validation, 0.4h unseen, 0.5h viVoice, 0.75h VIVOS
Manual sampling choices in Sections 2.1 and 3.2; small test sets limit statistical power for objective metrics.
Dialect labeling sample count = 20 audios per speaker
Hand-chosen in Appendix A for VALL-E dialect assignment; affects phonemization for each speaker.

assumptions (5)

domain assumption Whisper-large-v3 and PhoWhisper-large exact transcript match indicates correct transcription.
Section 2.1 filters mismatched transcripts; no human verification of the 940-hour training portion.
domain assumption wav2vec2-bartpho reliably detects multi-speaker audio segments.
Section 2.1 filters multi-speaker audio; no accuracy or failure analysis is reported.
domain assumption Speaker identities from audiobook metadata are reliable.
Each audiobook is assumed to have one narrator, yielding 735 distinct speakers; the paper does not validate this.
domain assumption PhoWhisper-large is a suitable reference ASR for WER evaluation.
Section 3.2 uses it to transcribe synthesized speech, and it was also used in curation, creating potential bias.
domain assumption Demucs vocal separation and sox volume normalization preserve speech quality needed for TTS training.
Pipeline in Section 2.1 applies them to all training audio without measuring degradation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Zero-Shot Text-to-Speech for Vietnamese." pith.science (2026). https://pith.science/paper/F6PUKEMD

@misc{pith2026250601322,
  author       = {Pith},
  title        = {Pith review of: Zero-Shot Text-to-Speech for Vietnamese},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F6PUKEMD}},
  note         = {Machine review of arXiv:2506.01322}
}

read the original abstract

This paper introduces PhoAudiobook, a newly curated dataset comprising 941 hours of high-quality audio for Vietnamese text-to-speech. Using PhoAudiobook, we conduct experiments on three leading zero-shot TTS models: VALL-E, VoiceCraft, and XTTS-V2. Our findings demonstrate that PhoAudiobook consistently enhances model performance across various metrics. Moreover, VALL-E and VoiceCraft exhibit superior performance in synthesizing short sentences, highlighting their robustness in handling diverse linguistic contexts. We publicly release PhoAudiobook to facilitate further research and development in Vietnamese text-to-speech.

Figures

Figures reproduced from arXiv: 2506.01322 by the authors.

**Figure 2.** Duration distributions of datasets. Audio sam [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 29 canonical work pages

[1]

Sercan Arik, Jitong Chen, Kainan Peng, Wei Ping, and Yanqi Zhou. 2018. Neural Voice Cloning with a Few Samples . In Proceedings of NeurIPS

work page 2018
[2]

James Betker. 2023. Better speech synthesis through scaling

work page 2023
[3]

Edresson Casanova, Kelly Davis, Eren Gölge, Görkem Göknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua Meyer, Reuben Morais, Samuel Olayemi, and Julian Weber. 2024. Xtts: a massively multilingual zero-shot text-to-speech model. In Proceedings of INTERSPEECH

work page 2024
[4]

Edresson Casanova, Julian Weber, Christopher D Shulby, Arnaldo Candido Junior, Eren G \"o lge, and Moacir A Ponti. 2022. Y our TTS : Towards Zero-Shot Multi-Speaker TTS and Zero-Shot Voice Conversion for Everyone . In Proceedings of ICML

work page 2022
[5]

Yuan-Jui Chen, Tao Tu, Cheng chieh Yeh, and Hung-Yi Lee. 2019. End-to-End Text-to-Speech for Low-Resource Languages by Cross-Lingual Transfer Learning . In Proceedings of INTERSPEECH

work page 2019
[6]

Erica Cooper, Cheng-I Lai, Yusuke Yasuda, Fuming Fang, Xin Wang, Nanxin Chen, and Junichi Yamagishi. 2020. Zero-Shot Multi-Speaker Text-To-Speech with State-Of-The-Art Neural Speaker Embeddings . In Proceedings of ICASSP

work page 2020
[7]

Alexandre D \'e fossez. 2021. Hybrid Spectrogram and Waveform Source Separation . In Proceedings of the ISMIR 2021 Workshop on Music Source Separation

work page 2021
[8]

Thinh Le Phuoc Gia, Tuan Pham Minh, Hung Nguyen Quoc, Trung Nguyen Quoc, and Vinh Truong Hoang. 2024. https://huggingface.co/datasets/capleaf/viVoice viVoice: Enabling Vietnamese Multi-Speaker Speech Synthesis

work page 2024

Show all 31 references

[9]

Alexander Gutkin, Linne Ha, Martin Jansche, Knot Pipatsrisawat, and Richard Sproat. 2016. TTS for Low Resource Languages: A B angla Synthesizer . In Proceedings of LREC

2016
[10]

Rongjie Huang, Chunlei Zhang, Yongqi Wang, Dongchao Yang, Jinchuan Tian, Zhenhui Ye, Luping Liu, Zehan Wang, Ziyue Jiang, Xuankai Chang, Jiatong Shi, Chao Weng, Zhou Zhao, and Dong Yu. 2024. Make-A-Voice: Revisiting Voice Large Language Models as Scalable Multilingual and Mult...

2024
[11]

Jaehyeon Kim, Jungil Kong, and Juhee Son. 2021. Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech . In Proceedings of ICML

2021
[12]

Thanh-Thien Le, Linh The Nguyen, and Dat Quoc Nguyen. 2024. PhoWhisper: Automatic Speech Recognition for Vietnamese . In Proceedings of the ICLR 2024 Tiny Papers track

2024
[13]

Jiaqi Li, Xueyao Zhang, Yuancheng Wang, Haorui He, Chaoren Wang, Li Wang, Huan Liao, Junyi Ao, Zeyu Xie, Yiqiao Huang, Junan Zhang, and Zhizheng Wu. 2025. Overview of the Amphion Toolkit (v0.2) . arXiv preprint, arXiv:2501.15442

2025 arXiv
[14]

Naihan Li, Shujie Liu, Yanqing Liu, Sheng Zhao, and Ming Liu. 2019. Neural speech synthesis with transformer network . In Proceedings of AAAI

2019
[15]

Yanqing Liu, Ruiqing Xue, Lei He, Xu Tan, and Sheng Zhao. 2022. DelightfulTTS 2: End-to-End Speech Synthesis with Adversarial Vector-Quantized Auto-Encoders . In Proceedings of INTERSPEECH

2022
[16]

Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual Denoising Pre-training for Neural Machine Translation . Transactions of the Association for Computational Linguistics, 8:726--742

2020
[17]

Hieu-Thi Luong and Hai-Quan Vu. 2016. A non-expert K aldi recipe for V ietnamese Speech Recognition System . In Proceedings of WLSI / OIAF 4 HLT , pages 51--55

2016
[18]

Florian Lux, Julia Koch, and Ngoc Thang Vu. 2022. Low-Resource Multilingual and Zero-Shot Multispeaker TTS . In Proceedings of AACL-IJCNLP

2022
[19]

Phuong Pham Ngoc, Chung Tran Quang, and Mai Luong Chi. 2023. ADAPT-TTS: HIGH-QUALITY ZERO-SHOT MULTI-SPEAKER TEXT-TO-SPEECH ADAPTIVE-BASED FOR VIETNAMESE . Journal of Computer Science and Cybernetics, 39(2):159–173

2023
[20]

Puyuan Peng, Po-Yao Huang, Shang-Wen Li, Abdelrahman Mohamed, and David Harwath. 2024. V oice C raft: Zero-Shot Speech Editing and Text-to-Speech in the Wild . In Proceedings of ACL

2024
[21]

Anh Pham, Khanh Linh Tran, Linh Nguyen, Thanh Duy Cao, Phuc Phan, and Duong A. Nguyen. 2024. https://github.com/quocanh34/Bud500 Bud500: A Comprehensive Vietnamese ASR Dataset

2024
[22]

Viet Thanh Pham, Xuan Thai Hoa Nguyen, Vu Hoang, and Thi Thu Trang Nguyen. 2023. Vietnam-Celeb: a large-scale dataset for Vietnamese speaker recognition . In Proceedings of INTERSPEECH

2023
[23]

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 ICML

2023
[24]

Mirco Ravanelli, Titouan Parcollet, Adel Moumen, Sylvain de Langen, Cem Subakan, Peter Plantinga, Yingzhi Wang, Pooneh Mousavi, Luca Della Libera, Artem Ploujnikov, Francesco Paissan, Davide Borra, Salah Zaiem, Zeyu Zhao, Shucong Zhang, Georgios Karakasidis, Sung-Lin Yeh, Pier...

2024
[25]

Yi Ren, Yangjun Ruan, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2019. FastSpeech: Fast, Robust and Controllable Text to Speech . In Proceedings of NeurIPS

2019
[26]

Simon Rouard, Francisco Massa, and Alexandre D \'e fossez. 2023. Hybrid Transformers for Music Source Separation . In Proceedings of ICASSP

2023
[27]

Xu Tan, Jiawei Chen, Haohe Liu, Jian Cong, Chen Zhang, Yanqing Liu, Xi Wang, Yichong Leng, Yuanhao Yi, Lei He, Sheng Zhao, Tao Qin, Frank Soong, and Tie-Yan Liu. 2024. NaturalSpeech: End-to-End Text-to-Speech Synthesis With Human-Level Quality . IEEE Transactions on Pattern An...

2024
[28]

VinBigData . 2023. https://institute.vinbigdata.org/en/events/vinbigdata-shares-100-hour-data-for-the-community VinBigData Shares 100-Hour Data for the Community

2023
[29]

Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, and Furu Wei. 2023. Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers . arXiv preprint, arXiv:2301.02111

2023 arXiv
[30]

Tao Wang, Jianhua Tao, Ruibo Fu, Jiangyan Yi, Zhengqi Wen, and Rongxiu Zhong. 2020. Spoken Content and Voice Factorization for Few-Shot Speaker Adaptation . In Proceedings of INTERSPEECH

2020
[31]

Yihan Wu, Xu Tan, Bohan Li, Lei He, Sheng Zhao, Ruihua Song, Tao Qin, and Tie-Yan Liu. 2022. AdaSpeech 4: Adaptive Text to Speech in Zero-Shot Scenarios . In Proceedings of INTERSPEECH

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Sercan Arik, Jitong Chen, Kainan Peng, Wei Ping, and Yanqi Zhou. 2018. Neural Voice Cloning with a Few Samples . In Proceedings of NeurIPS

work page 2018

[2] [2]

James Betker. 2023. Better speech synthesis through scaling

work page 2023

[3] [3]

Edresson Casanova, Kelly Davis, Eren Gölge, Görkem Göknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua Meyer, Reuben Morais, Samuel Olayemi, and Julian Weber. 2024. Xtts: a massively multilingual zero-shot text-to-speech model. In Proceedings of INTERSPEECH

work page 2024

[4] [4]

Edresson Casanova, Julian Weber, Christopher D Shulby, Arnaldo Candido Junior, Eren G \"o lge, and Moacir A Ponti. 2022. Y our TTS : Towards Zero-Shot Multi-Speaker TTS and Zero-Shot Voice Conversion for Everyone . In Proceedings of ICML

work page 2022

[5] [5]

Yuan-Jui Chen, Tao Tu, Cheng chieh Yeh, and Hung-Yi Lee. 2019. End-to-End Text-to-Speech for Low-Resource Languages by Cross-Lingual Transfer Learning . In Proceedings of INTERSPEECH

work page 2019

[6] [6]

Erica Cooper, Cheng-I Lai, Yusuke Yasuda, Fuming Fang, Xin Wang, Nanxin Chen, and Junichi Yamagishi. 2020. Zero-Shot Multi-Speaker Text-To-Speech with State-Of-The-Art Neural Speaker Embeddings . In Proceedings of ICASSP

work page 2020

[7] [7]

Alexandre D \'e fossez. 2021. Hybrid Spectrogram and Waveform Source Separation . In Proceedings of the ISMIR 2021 Workshop on Music Source Separation

work page 2021

[8] [8]

Thinh Le Phuoc Gia, Tuan Pham Minh, Hung Nguyen Quoc, Trung Nguyen Quoc, and Vinh Truong Hoang. 2024. https://huggingface.co/datasets/capleaf/viVoice viVoice: Enabling Vietnamese Multi-Speaker Speech Synthesis

work page 2024