Pith. sign in

REVIEW 4 major objections 5 minor 9 cited by

IndexTTS: An Industrial-Level Controllable and Efficient Zero-Shot Text-To-Speech System

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

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

Pith's one-line read IndexTTS is a zero-shot text-to-speech system that, according to the paper, surpasses existing open-source TTS systems in naturalness, content consistency, and voice-cloning similarity by combining a hybrid character-pinyin input scheme…

desk verdict A credible industrial zero-shot TTS system with a genuinely useful polyphone-control mechanism, but the abstract's faster-inference claim is contradicted by the paper's own Table 5. read the letter →

arxiv 2502.05512 v1 pith:YTAC3FXO submitted 2025-02-08 cs.SD cs.AIeess.AS

Wei Deng , Siyi Zhou , Jingchen Shu , Jinchao Wang , Lu Wang This is my paper · ORCID
classification cs.SDcs.AIeess.AS
keywords zero-shottext-to-speechvoicecloningcharacter-pinyinhybridmodelingpolyphonedisambiguationfinitescalarquantizationBigVGAN2codebookutilizationspeechlanguagemodel
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 IndexTTS, a zero-shot text-to-speech system built on the XTTS and Tortoise architectures. It argues that by combining a hybrid Chinese-character-and-pinyin input scheme with a Conformer-based speaker conditioning encoder and a BigVGAN2 decoder, the system achieves better naturalness, content consistency, and voice-cloning similarity than leading open-source TTS systems such as XTTS, CosyVoice2, Fish-Speech, FireRedTTS, and F5-TTS. The authors also report that switching from VQ to FSQ nearly eliminates codebook collapse in the speech codec, and that pinyin input corrects 94% of polyphonic-character pronunciation errors in their test set. If the claims hold, IndexTTS offers a practical combination of controllability, speed, and cloning fidelity for industrial use.

What carries the argument

The system is built around three interacting components: a hybrid character-pinyin BPE tokenizer that makes Chinese pronunciation controllable, a Conformer-based Perceiver conditioner that compresses prompt speech into speaker vectors for stable zero-shot cloning, and a BigVGAN2 decoder that reconstructs the 24 kHz waveform directly from the LLM's last hidden states. The speech codec uses VQ with an 8,192-code codebook at a 25 Hz token rate, and the paper compares it with FSQ to show near-100% codebook utilization at 34k training hours. This machinery carries the argument by converting large-scale pseudo-labeled data into improvements on content consistency, speaker similarity, and prosody.

What would settle it

Reproduce IndexTTS from the released checkpoints and test sets, and run the same Aishell-1, CommonVoice, and LibriSpeech evaluations with the identical ASR and speaker-embedding models; if an independent run does not approach the reported averaged WER of 3.7% and speaker similarity of 0.776, or if a neutral benchmark shows XTTS or CosyVoice2 matching IndexTTS on content consistency and similarity, the central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a GPT-style text-to-codec language model, conditioned by a Conformer-based Perceiver encoder on a prompt speech segment and trained on 34,000 hours of ASR-labeled bilingual audio, can simultaneously deliver state-of-the-art zero-shot voice cloning and content consistency while allowing direct user correction of Chinese polyphonic character pronunciation through pinyin. The system's architecture—removing the front-end G2P module, tokenizing raw text with BPE, mixing pinyin into training data, and decoding with BigVGAN2—produces averaged WER of 3.7% and speaker similarity of 0.776 across four test sets, outperforming all compared open-source systems on both objective and subjective metrics while using less GPU time than most.

Load-bearing premise

The whole comparison rests on the assumption that the 34,000 hours of internet audio, labeled automatically by ASR with added punctuation, carry accurate enough transcripts that the model learns correct pronunciations and the evaluation's WER numbers reflect genuine content fidelity rather than label errors.

Editorial extensions

If this is right

  • IndexTTS achieves lower averaged WER (3.7%) and higher speaker similarity (0.776) than XTTS, CosyVoice2, Fish-Speech, FireRedTTS, and F5-TTS on the four evaluated test sets.
  • The character-pinyin hybrid input lets users correct polyphonic character pronunciations directly, fixing 94% of the errors found when using characters alone.
  • Replacing VQ with FSQ nearly eliminates codebook collapse at 6k hours of training data, while at 34k hours VQ also approaches 100% utilization, showing that data scale mitigates codebook collapse.
  • The conformer-based conditioning encoder and BigVGAN2 decoder improve timbre similarity and training stability, contributing to an average MOS of 4.01.
  • Because inference needs no prompt text, cross-lingual voice cloning is feasible and the system achieves faster synthesis (397 seconds for 200 samples at 28.47% GPU utilization) than most baselines.

Reading between the lines

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

  • The 94% pinyin-correction rate suggests a practical user interface where users can spot-fix pronunciations without retraining, a pattern that could generalize to other languages with homographs beyond Chinese.
  • If pseudo-label noise in ASR transcripts is the main error source, then scaling data or filtering labels by confidence could further push the 1.1% uncorrectable polyphone errors toward zero.
  • The finding that VQ's codebook utilization deficit disappears with 34k hours questions the need for FSQ in large-scale TTS and may inform codec design at different data scales.
  • The reported GPU utilization of 28.47% for IndexTTS is much lower than baselines, implying headroom for faster-than-real-time synthesis on modest hardware, though this is the authors' measurement, not an independent benchmark.
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 IndexTTS, a zero-shot TTS system built on an XTTS/Tortoise-style architecture with three main modifications: a character-pinyin hybrid tokenizer for Chinese pronunciation control, a Conformer-based conditioning encoder, and a BigVGAN2-based speech decoder; it also compares VQ and FSQ speech codecs. The experiments report WER and speaker similarity on LibriSpeech, Aishell-1, and CommonVoice subsets, a 2,500-item polyphonic-character test, MOS listening results, and inference timing. The authors claim that IndexTTS surpasses XTTS, CosyVoice2, Fish-Speech, FireRedTTS, and F5-TTS in both quality and speed.

Significance. The engineering contribution is potentially useful: the released test sets, including the polyphonic-character test set, the near-100% codebook utilization result, and the 94% correction rate among mispronounced polyphonic characters are concrete strengths. The system is clearly described, and the limitations section is honest about missing emotion and language coverage. However, the central comparative claims are not yet backed by the reported evidence: the timing data contradict the speed claim, the speaker-similarity numbers do not consistently favor IndexTTS, and no uncertainty quantification is provided. These issues are fixable with additional analysis and revised claims, so the contribution is viable but currently overstated.

major comments (4)
  1. [Abstract; §3.3.3, Table 5] The claim that IndexTTS has 'faster inference speed' than F5-TTS is contradicted by the paper's own Table 5: on the same 200-sample benchmark, IndexTTS takes 397 s while F5-TTS takes 320 s. The table reports no hardware, batch size, number of runs, or whether the duration is end-to-end or model-only, so the GPU-utilization column cannot rescue the speed claim. Please correct or substantially qualify the speed claim and report a complete timing protocol.
  2. [§3.3.3, Table 3] The claim that IndexTTS 'significantly outperforms' all baselines is not supported by the speaker-similarity numbers: in the AVG column, IndexTTS has 0.776 versus CosyVoice2's 0.788 and F5-TTS's 0.779, and on Aishell-1 and CommonVoice-zh its SS is below CosyVoice2 (0.744 vs 0.796) and F5-TTS (0.742 vs 0.747), respectively. The WER results are strong, but the paper needs a clear aggregate criterion or per-set significance tests before claiming overall superiority.
  3. [§3.3.3, Table 4] The MOS result is based on 100 randomly selected samples with no confidence intervals, no number of raters, and no significance test. The reported differences (e.g., prosody 3.79 vs 3.67 for CosyVoice2, quality 4.05 vs 3.73) may be meaningful, but without uncertainty quantification the claim of 'significant advantages in timbre similarity and sound quality' is not established.
  4. [§3.1; §3.2.3] The evaluation is vulnerable to underexplained configuration choices. Please describe the exact baseline checkpoints, decoding parameters, prompt-selection procedure, ASR label accuracy, and punctuation-insertion heuristics. This matters because the system was trained on 34,000 hours of ASR pseudo-labels and the paper itself attributes 28 residual polyphonic errors to training-data errors; a systematic label-noise estimate would help assess whether the comparisons are fair.
minor comments (5)
  1. [Table 2] The 94.0% in row A2 should be explicitly defined as the fraction of the 465 A1 errors that are corrected; as printed, the denominator is ambiguous.
  2. [References, [14]] Reference [14] is cited for BigVGAN2, but the reference title is 'BigVGAN: A universal neural vocoder...'; please cite the BigVGAN2 paper specifically.
  3. [§2.3] The text moves from 'Conformer encoder' to 'Conformer-based Perceiver' without introducing the Perceiver module; clarify the relationship and define the perceptual component of the conditioning encoder.
  4. [Figure 1] The label 'Acusitic Token' should be 'Acoustic Token'.
  5. [§3.3.3] The sentence 'we have outperformed the baseline in nearly all evaluation dimensions' should read 'the baselines' and should specify which dimension was not outperformed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are supported by forward evaluations on external benchmarks and held-out test sets, not by fitted parameters or load-bearing self-citations.

full rationale

The paper's claimed derivation chain is architectural and empirical rather than definitional. The central comparisons in Tables 3, 4, and 5 use standard external corpora (LibriSpeech, Aishell-1, CommonVoice), a separately constructed polyphonic-character test set, and MOS listening tests; no reported quantity is reconstructed from a fitted parameter or from an equation that contains its own output. The 34,000-hour ASR-pseudo-labeled training set is an input condition, and the WER/SS numbers are forward measurements on different test utterances, so the evaluation does not reduce to the training labels. There are no author self-citations used as load-bearing support, no imported uniqueness theorem, and no ansatz smuggled in via citation. The pinyin-mixing training scheme is an intervention whose correction rate is measured on a dedicated 2,500-sentence test set, not asserted by construction. The one notable defect in the paper is an internal inconsistency, not circularity: Table 5 reports F5-TTS completing the 200-sample benchmark in 320 seconds versus IndexTTS's 397 seconds, which contradicts the abstract's blanket 'faster inference speed' claim; GPU utilization is also reported without hardware or batch-size context. That is a correctness/evidence concern under the reviewer's rules, not a circular reduction, and it does not change the circularity verdict. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper's central claims rely on several hand-chosen design hyperparameters and on the assumption that noisy ASR labels and automatic speaker embeddings are adequate ground truth for training and evaluation. No new physical or abstract entities are introduced.

free parameters (3)
  • Pinyin replacement proportions = 50% of samples, 20% of characters
    Chosen by hand without ablation; the training mixes pinyin into the text at these rates.
  • FSQ levels = [8, 8, 8, 6, 5]
    Hand-picked configuration for the scalar quantizer; no search or justification is given.
  • Conformer subsample rate = 2
    Selected for the conditioning encoder; no comparison to other rates.
assumptions (3)
  • domain assumption ASR pseudo-labels on 34,000 hours of internet audio are accurate enough to serve as training text.
    Section 3.1 states the dataset uses ASR-generated pseudo-labels; no analysis of label noise is provided.
  • domain assumption Speaker similarity measured by ERes2Net2 cosine similarity reflects perceptual voice cloning quality.
    Section 3.2.3 describes the metric; the paper does not validate it against human judgments.
  • domain assumption The random prompt selection protocol gives a fair comparison across systems.
    Section 3.2.3 selects one other sample as prompt; no check that prompt difficulty is balanced across systems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IndexTTS: An Industrial-Level Controllable and Efficient Zero-Shot Text-To-Speech System." pith.science (2026). https://pith.science/paper/YTAC3FXO

@misc{pith2026250205512,
  author       = {Pith},
  title        = {Pith review of: IndexTTS: An Industrial-Level Controllable and Efficient Zero-Shot Text-To-Speech System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YTAC3FXO}},
  note         = {Machine review of arXiv:2502.05512}
}
read the original abstract

Recently, large language model (LLM) based text-to-speech (TTS) systems have gradually become the mainstream in the industry due to their high naturalness and powerful zero-shot voice cloning capabilities.Here, we introduce the IndexTTS system, which is mainly based on the XTTS and Tortoise model. We add some novel improvements. Specifically, in Chinese scenarios, we adopt a hybrid modeling method that combines characters and pinyin, making the pronunciations of polyphonic characters and long-tail characters controllable. We also performed a comparative analysis of the Vector Quantization (VQ) with Finite-Scalar Quantization (FSQ) for codebook utilization of acoustic speech tokens. To further enhance the effect and stability of voice cloning, we introduce a conformer-based speech conditional encoder and replace the speechcode decoder with BigVGAN2. Compared with XTTS, it has achieved significant improvements in naturalness, content consistency, and zero-shot voice cloning. As for the popular TTS systems in the open-source, such as Fish-Speech, CosyVoice2, FireRedTTS and F5-TTS, IndexTTS has a relatively simple training process, more controllable usage, and faster inference speed. Moreover, its performance surpasses that of these systems. Our demos are available at https://index-tts.github.io.

Figures

Figures reproduced from arXiv: 2502.05512 by the authors.

Figure 1
Figure 1. An overview of IndexTTS, a text-to-speech language model conditioned on prompt speech and text tokens generates acoustic tokens, and the BigVGAN2 decoder convert the LLM output latent into waveform [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Compare the distribution of codebook utilization rates of VQ and FSQ under different training data scales 3.3.3. Comparison Results with Baselines We select several top popular zero-shot TTS models in the opensource for comparison, including systems XTTS[1], CosyVoice2(non-streaming)[22], FishSpeech[2], FireRedTTS[4] and F5-TTS[5]. The evaluation methodology encompasses both objective and subjective metrics: the wor… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 9 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. AST: Adaptive, Seamless, and Training-Free Precise Speech Editing

    cs.SD 2026-04 unverdicted novelty 7.0 of 10

    AST performs training-free text-based speech editing by stitching inverted source latents with synthesized targets and adaptively guiding the flow-matching decoder, achieving state-of-the-art temporal fidelity and spe...

  2. DiTReducio: A Training-Free Acceleration for DiT-Based TTS via Progressive Calibration

    cs.SD 2025-09 conditional novelty 6.0 of 10

    DiTReducio is a training-free, pattern-guided layer and branch skipping method that accelerates DiT-based TTS, reporting significant FLOP and RTF reductions with modest quality loss at tuned thresholds.

  3. X-Translator: A Real-Time Multilingual Speaker-Aware Speech-to-Speech Translation System

    eess.AS 2026-07 conditional novelty 5.0 of 10

    An open, modular cascaded system (streaming ASR + MT + prompt-conditioned TTS) preserves speaker identity in long-form multi-speaker translation, at higher latency and slightly lower translation quality than proprietary APIs.

  4. AutoSIFT: Automatic Style Sifting for Controllable Speech Generation with Arbitrary Style Infilling

    cs.SD 2026-07 unverdicted novelty 5.0 of 10

    AutoSIFT disentangles text-describable style categories from residual speech styles and selectively infills only the categories the user specifies.

  5. FireRedTTS-2: Towards Long Conversational Speech Generation for Podcast and Chatbot

    cs.SD 2025-09 conditional novelty 5.0 of 10

    FireRedTTS-2 generates long multi-speaker conversations in a streaming, sentence-by-sentence way using a new low-rate speech tokenizer and a dual-transformer text-speech model.

  6. IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech

    cs.CL 2025-06 conditional novelty 5.0 of 10

    IndexTTS2 achieves precise token-count-based duration control and emotion/speaker disentanglement in an autoregressive zero-shot TTS, reporting SOTA WER, speaker similarity, and emotional fidelity.

  7. CloneShield: A Framework for Universal Perturbation Against Zero-Shot Voice Cloning

    cs.SD 2025-05 reject novelty 5.0 of 10

    A universal adversarial perturbation framework claiming to protect speech against zero-shot voice cloning by degrading cloned outputs while preserving input naturalness.

  8. UniTTS: An end-to-end TTS system without decoupling of acoustic and semantic information

    cs.SD 2025-05 conditional novelty 5.0 of 10

    The authors propose DistilCodec, a 32,768-code single-codebook audio codec, and UniTTS, a Qwen2.5-7B TTS model trained with audio, text, and cross-modal autoregressive tasks on interleaved prompts.

  9. Robust Residual Finite Scalar Quantization for Neural Compression

    eess.IV 2025-08 reject novelty 4.0 of 10

    RFSQ applies learned scaling or invertible LayerNorm to residual FSQ to prevent magnitude decay, reporting DNSMOS and image loss gains, though the LayerNorm variant has a reconstruction inconsistency.

Reference graph

Works this paper leans on

25 extracted references · 2 canonical work pages · cited by 9 Pith papers

  1. [1]

    Introduction Text-to-speech synthesis (TTS) has extensive applications in fields such as human-computer interaction, education, and en- tertainment. For example, in video creation scenarios in recent years, TTS can assist users in quickly generating video dubbing, saving recording time, and thus playing a crucial role in the cre- ation process. Many creat...

    arXiv 2025

  2. [2]

    [BT], prompt text, text, [ET], [BA], prompt audio, audio, [EA]

    IndexTTS System Similar to XTTS[1], our system incorporates speech-to-codec VQV AE[12] codec, text-to-codec language model and latent- to-audio decoder, as depicted in Figure 1. 2.1. Text tokenizer Currently, our system only supports two languages, Chinese and English. We directly use the raw text as input, which is tokenized by a BPE-based text tokenizer...

    work page

  3. [3]

    Dataset All training data was collected from the internet, with an initial 120,000 hours of raw audio

    Experiments 3.1. Dataset All training data was collected from the internet, with an initial 120,000 hours of raw audio. After voice separation, speaker segmentation, and filtering using Demucs [16], we obtained 34,000 hours of high-quality Chinese-English bilingual data. The dataset includes 25,000 hours of Chinese and 9,000 hours of English audio. We the...

    work page

  4. [4]

    Xtts: a massively multilingual zero-shot text-to-speech model,

    E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea, L. Hart, A. Aljafari, J. Meyer, R. Morais, S. Olayemi et al. , “Xtts: a massively multilingual zero-shot text-to-speech model,” arXiv preprint arXiv:2406.04904, 2024

    arXiv 2024

  5. [5]

    Fish-speech: Leveraging large language models for advanced multilingual text-to-speech synthesis,

    S. Liao, Y . Wang, T. Li, Y . Cheng, R. Zhang, R. Zhou, and Y . Xing, “Fish-speech: Leveraging large language models for advanced multilingual text-to-speech synthesis,” arXiv preprint arXiv:2411.01156, 2024

    arXiv 2024

  6. [6]

    Cosyvoice: A scalable multi- lingual zero-shot text-to-speech synthesizer based on supervised semantic tokens,

    Z. Du, Q. Chen, S. Zhang, K. Hu, H. Lu, Y . Yang, H. Hu, S. Zheng, Y . Gu, Z. Ma et al. , “Cosyvoice: A scalable multi- lingual zero-shot text-to-speech synthesizer based on supervised semantic tokens,” arXiv preprint arXiv:2407.05407, 2024

    arXiv 2024

  7. [7]

    Fireredtts: A foundation text-to-speech framework for industry-level generative speech applications,

    H.-H. Guo, K. Liu, F.-Y . Shen, Y .-C. Wu, F.-L. Xie, K. Xie, and K.-T. Xu, “Fireredtts: A foundation text-to-speech framework for industry-level generative speech applications,” arXiv preprint arXiv:2409.03283, 2024

    arXiv 2024

  8. [8]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,” arXiv preprint arXiv:2410.06885, 2024

    arXiv 2024

Show all 25 references
  1. [9]

    Mega-tts 2: Zero-shot text- to-speech with arbitrary length speech prompts,

    Z. Jiang, J. Liu, Y . Ren, J. He, C. Zhang, Z. Ye, P. Wei, C. Wang, X. Yin, Z. Ma et al. , “Mega-tts 2: Zero-shot text- to-speech with arbitrary length speech prompts,” arXiv preprint arXiv:2307.07218, 2023

    2023 arXiv

  2. [10]

    Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,

    E. Casanova, J. Weber, C. D. Shulby, A. C. Junior, E. G ¨olge, and M. A. Ponti, “Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,” in International Con- ference on Machine Learning. PMLR, 2022, pp. 2709–2720

    2022

  3. [11]

    Neural codec language mod- els are zero-shot text to speech synthesizers,

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li et al. , “Neural codec language mod- els are zero-shot text to speech synthesizers,” arXiv preprint arXiv:2301.02111, 2023

    2023 arXiv

  4. [12]

    Seed-tts: A family of high-quality versatile speech generation models,

    P. Anastassiou, J. Chen, J. Chen, Y . Chen, Z. Chen, Z. Chen, J. Cong, L. Deng, C. Ding, L. Gao et al., “Seed-tts: A family of high-quality versatile speech generation models,” arXiv preprint arXiv:2406.02430, 2024

    2024 arXiv

  5. [13]

    Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,” Advances in neural information processing systems , vol. 33, pp. 17 022– 17 033, 2020

    2020

  6. [14]

    Better speech synthesis through scaling,

    J. Betker, “Better speech synthesis through scaling,” arXiv preprint arXiv:2305.07243, 2023

    2023 arXiv

  7. [15]

    Neural discrete represen- tation learning,

    A. Van Den Oord, O. Vinyals et al. , “Neural discrete represen- tation learning,” Advances in neural information processing sys- tems, vol. 30, 2017

    2017

  8. [16]

    Fi- nite scalar quantization: Vq-vae made simple,

    F. Mentzer, D. Minnen, E. Agustsson, and M. Tschannen, “Fi- nite scalar quantization: Vq-vae made simple,” arXiv preprint arXiv:2309.15505, 2023

    2023 arXiv

  9. [17]

    Bigvgan: A universal neural vocoder with large-scale training,

    S.-g. Lee, W. Ping, B. Ginsburg, B. Catanzaro, and S. Yoon, “Bigvgan: A universal neural vocoder with large-scale training,” arXiv preprint arXiv:2206.04658, 2022

    2022 arXiv

  10. [18]

    Matcha-tts: A fast tts architecture with conditional flow match- ing,

    S. Mehta, R. Tu, J. Beskow, ´E. Sz ´ekely, and G. E. Henter, “Matcha-tts: A fast tts architecture with conditional flow match- ing,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 341–11 345

    2024

  11. [19]

    Demucs: Deep extractor for music sources with extra unlabeled data remixed,

    A. D ´efossez, N. Usunier, L. Bottou, and F. Bach, “Demucs: Deep extractor for music sources with extra unlabeled data remixed,” arXiv preprint arXiv:1909.01174, 2019

    1909 arXiv

  12. [20]

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

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  13. [21]

    Aishell-1: An open- source mandarin speech corpus and a speech recognition base- line,

    H. Bu, J. Du, X. Na, B. Wu, and H. Zheng, “Aishell-1: An open- source mandarin speech corpus and a speech recognition base- line,” in 2017 20th conference of the oriental chapter of the inter- national coordinating committee on speech databases and speech I/O systems and asses...

    2017

  14. [22]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,”arXiv preprint arXiv:1912.06670, 2019

    1912 arXiv

  15. [23]

    Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to- end speech recognition,

    Z. Gao, S. Zhang, I. McLoughlin, and Z. Yan, “Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to- end speech recognition,” arXiv preprint arXiv:2206.08317, 2022

    2022 arXiv

  16. [24]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” in International conference on machine learning . PMLR, 2023, pp. 28 492–28 518

    2023

  17. [25]

    Cosyvoice 2: Scalable stream- ing speech synthesis with large language models,

    Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wanget al., “Cosyvoice 2: Scalable stream- ing speech synthesis with large language models,” arXiv preprint arXiv:2412.10117, 2024

    2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.