Pith. sign in

REVIEW 3 major objections 5 minor 18 references

Quantize More, Lose Less: Autoregressive Generation from Residually Quantized Speech Representations

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A multi-codebook audio codec with ASR-guided first-layer quantization can make text-to-speech more expressive than single-codebook systems.

desk verdict Reasonable multi-codebook TTS architecture, but the paper's central superiority claim is contradicted by its own Table 6, where QTTS-8 loses to CosyVoice2 on WER, speaker similarity, and PGChard, and the touted QTTS-16 model does not appear in the synthesis table. read the letter →

arxiv 2507.12197 v1 pith:2Z4JPQFB submitted 2025-07-16 cs.SD cs.AI

Yichen Han , Xiaoyang Hao , Keming Chen , Weibo Xiong , Jun He , Ruonan Zhang , Junjie Cao , Yue Liu
show 20 more authors
Bowen Li Dongrui Zhang Hui Xia Huilei Fu Kai Jia Kaixuan Guo Mingli Jin Qingyun Meng Ruidong Ma Ruiqian Fang Shaotong Guo Xuhui Li Yang Xiang Ying Zhang Yulong Liu Yunfeng Li Yuyi Zhang Yuze Zhou Zhen Wang Zhaowen Chen
This is my paper
classification cs.SDcs.AI
keywords text-to-speechresidualvectorquantizationaudiocodecautoregressivetransformersemanticdisentanglementzero-shotTTSneuralcompressionRVQ
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 is trying to establish that the reason discrete text-to-speech models lose expressive detail is the single-codebook bottleneck, and that multi-codebook residual quantization can fix it if the codebooks are forced to separate semantic content from acoustic detail. It proposes a codec whose first codebook is supervised by an ASR loss and whose later codebooks are left to capture residuals, plus two decoding strategies that make multi-codebook generation practical. If the claim holds, scaling up compression by adding codebooks becomes a direct path to higher-fidelity, more expressive speech, and possibly to singing and music generation. The reported results put the multi-codebook system ahead of single-codebook baselines on objective and subjective measures.

What carries the argument

The central object is residual vector quantization with a semantic-supervision split. At each frame, the encoder vector $f_t$ is quantized in $C$ stages, where stage $i$ chooses the codeword $e_{c_t^{(i)}}^{(i)}$ closest to the residual $r_t^{(i-1)}$, subtracts it to form $r_t^{(i)}$, and the final reconstruction is $\hat f_t = \sum_{i=1}^C e_{c_t^{(i)}}^{(i)}$. The load-bearing twist is that an autoregressive ASR model reads only tokens from the first codebook and its cross-entropy loss is backpropagated only through that codebook, while the other codebooks are free to fit the residual acoustic detail under the reconstruction and adversarial losses. The two generation heads then handle the multi-codebook stream: the dual-autoregressive decoder predicts time then codebook depth to capture inter-codebook dependencies, and the delay-multihead decoder uses one prediction head per codebook with fixed delays so all heads emit tokens in parallel.

What would settle it

Train the proposed codec without the ASR cross-entropy term and compare word error rate and reconstruction quality: if the synthesized speech does not degrade materially on WER or speaker similarity, the semantic supervision is not responsible for the claimed gains and the disentanglement mechanism is not load-bearing.

Watch

Extended reading notes

Core claim

The paper's central claim is that a text-to-speech system built on multi-codebook residual vector quantization can outperform single-codebook systems when the codebooks are explicitly separated by information type. Its codec trains an ASR model on the first codebook's tokens alone, so the first codebook is pressured to carry linguistic content while the remaining codebooks carry residual acoustic detail; reconstruction then uses all codebooks. On top of this, two autoregressive decoders are proposed: a dual-autoregressive decoder that conditions each codebook on all previous ones for maximum fidelity, and a delay-multihead decoder that predicts all codebooks in parallel with a fixed step offset for speed. The experiments report that a 16-codebook configuration achieves near-lossless audio reconstruction and that the full system beats single-codebook baselines on word error rate, speaker similarity, and listening quality.

Load-bearing premise

The load-bearing premise is that pushing the ASR transcription loss only through the first codebook actually forces a clean split, so that codebook one holds the words and the other codebooks hold only voice quality, timing, and timbre.

Editorial extensions

If this is right

  • Scaling from 1 to 16 residual codebooks moves reconstruction from lossy to near-lossless on the reported metrics, raising PESQ and speaker similarity while lowering word error rate.
  • A multi-codebook system with 8 or 16 codebooks beats single-codebook baselines on all reported synthesis metrics, supporting the claim that the single-codebook bottleneck is the main source of lost expressiveness.
  • The dual-autoregressive decoder maximizes quality by preserving full inter-codebook dependencies, while the delay-multihead decoder trades some contextual visibility for much lower inference latency.
  • The fixed-delay design allows the decoder to be compiled into a static computation graph, which keeps time-to-first-token low and token throughput high even with many codebooks.
  • Because the codec reconstruction gains transfer to downstream synthesis, improving the codec is a reliable lever for improving overall text-to-speech quality.

Reading between the lines

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

  • A direct test the paper does not report is measuring how much linguistic content leaks into the residual codebooks: a phoneme classifier or mutual-information estimate on codebooks 2 through 8 would settle whether the ASR supervision actually achieves the claimed separation or is overridden by the reconstruction loss.
  • If the semantic/acoustic split holds, the same recipe could plausibly extend to singing and music synthesis by treating lyrics or pitch contours as the semantic stream and leaving timbre and production detail to the residual codebooks.
  • The delay-multihead design suggests that adding more codebooks need not add proportional autoregressive latency, since the parallel heads operate with a fixed delay offset rather than a full serial pass.
  • A natural next step, not tested here, would be adding an explicit orthogonality or adversarial penalty between codebook streams to guard against semantic leakage if the ASR loss alone proves insufficient.
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

3 major / 5 minor

Summary. The paper proposes QDAC, a residual vector quantization (RVQ) audio codec trained with a GAN-based reconstruction loss and an autoregressive ASR head that supervises only the first codebook, with the goal of disentangling semantic content from acoustic detail. On top of QDAC, the authors build QTTS, a TTS system with two decoding strategies: Hierarchical Parallel, which autoregressively models both time and codebook dimensions, and Multihead Delay, which parallelizes codebook prediction with a fixed offset. The paper reports codec reconstruction metrics, zero-shot TTS objective metrics, and inference-speed benchmarks, and claims that QTTS substantially surpasses single-codebook baselines on all metrics.

Significance. If the central claims were fully supported, the paper would make a useful contribution: it addresses a recognized limitation of single-codebook TTS and proposes a concrete multi-codebook architecture with two explicit quality/speed trade-offs. The codec reconstruction results in Table 5 (e.g., QDAC-16 at 50 Hz improving PESQ and SI-SDR over lower-codebook configurations) and the detailed inference-throughput measurements are potentially valuable empirical data points. However, the main synthesis-quality claim is currently contradicted by the paper's own Table 6, and the semantic-disentanglement mechanism is asserted without supporting analysis. These issues must be resolved before the contribution can be assessed.

major comments (3)
  1. [Section 3.2, Table 6] The text immediately after Table 6 states that the '16-codebook model substantially surpasses the single-codebook baselines (CosyVoice v1/v2) on all metrics,' but Table 6 reports only an 8-codebook QTTS model. In that table, QTTS-8 has worse WER (1.66 vs. 1.32), worse PGChard (6.89 vs. 6.11), and worse speaker similarity (0.75 vs. 0.81) than CosyVoice2; only MOS is marginally higher (3.03 vs. 3.02). The displayed evidence therefore contradicts the dominance claim, and the claimed 16-codebook synthesis result is absent from the table. Please provide the 16-codebook synthesis results or revise the claim to match the presented data.
  2. [Section 2.1.2] The core mechanism of QDAC is the claim that backpropagating the ASR cross-entropy only through C1 forces C1 to encode semantic content while C2 through CN capture purely residual acoustic detail. This assumption is load-bearing for every downstream quality and WER advantage claimed for QDAC, but the paper provides no ablation (e.g., training without the ASR loss), no codebook-usage statistics, and no mutual-information or error analysis demonstrating that semantic information is confined to C1 and does not leak into the other codebooks. Without such evidence, the 'disentanglement' is an assumption rather than a demonstrated property.
  3. [Table 5 and Section 3.2 setup] WER is reported as a codec quality metric, but because the QDAC training objective includes the ASR cross-entropy loss on C1, low WER partly reflects the training objective and is not an independent measure of semantic fidelity. In addition, Table 5 has a corrupted bitrate entry ('4.4bos') and a missing bitrate for the first QDAC row, and the setup text says '1, 4, 8, and 16 codebooks' were evaluated while the table only shows 8 and 16 codebooks. These issues make the codec evaluation difficult to verify and the stated trend only partially supported.
minor comments (5)
  1. [Section 2.1.2] There is a typo 'As show in Figure 1' that should read 'As shown in Figure 1'; likewise, 'delay-parttern' should be 'delay-pattern'.
  2. [Table 2] The rows of Table 2 are incorrectly merged (for example, '228256' appears to combine several values across columns), making the latency results unreadable and unverifiable.
  3. [Section 3.2 setup] The sentence describing the codebook configurations says '1, 4, 8, and 16 codebooks,' but Table 5 lists only 8 and 16 codebooks; either supply the omitted configurations or correct the text.
  4. [Table 6] The abbreviation 'PGChard' is not defined anywhere in the paper, and the direction of improvement (lower is better) is not stated; please define it and explain the metric.
  5. [Section 3.1] The text refers to 'Table 3.1' and 'Table 3.1' in the inference-acceleration subsection, but the actual tables are numbered 1 through 4; this cross-reference needs to be corrected.

Circularity Check

1 steps flagged · score 5.0 of 10

WER is trained into QDAC via an ASR loss on C1 and then reported as evidence of semantic fidelity, a genuine but partial metric-level circularity; the Table 6 evidence mismatch is a correctness defect, not circularity.

  1. fitted input called prediction [Section 2.1.2 (training objective) versus Table 5 and Abstract (evaluation claims)]
    "The cross-entropy loss from this ASR prediction is therefore backpropagated only through this primary codebook. This acts as a powerful supervisory signal, forcing C1 to encode the semantic content necessary for transcription, while leaving the other codebooks C2...N free to be optimized by the reconstruction loss for capturing non-semantic information like speaker timbre and prosody."

    Section 2.1.2 trains QDAC end-to-end so that C1 alone suffices for ASR transcription ('forcing C1 to encode the semantic content necessary for transcription'). Table 5 then reports WER on decoded audio as a codec-quality metric, and the Abstract claims QDAC 'better preserves expressive content.' The reported QDAC WER of 5.22 is below the paper's own ground-truth WER of 6.01, which is only explainable if the WER metric is directly optimized by the ASR objective rather than measuring independent semantic preservation. Reporting as evidence of quality a quantity that is the evaluation image of the training loss is the fitted-input-called-prediction pattern: the low WER is statistically forced by construction.

full rationale

Derivation-chain walk: (1) QDAC sets up cascaded RVQ; (2) Section 2.1.2 adds an ASR cross-entropy loss backpropagated only through C1 and asserts, without derivation, that this forces C1 to be linguistic while C2...N capture residual acoustics; (3) QTTS decodes QDAC tokens with the two AR strategies; (4) Section 3.2 evaluates codec reconstruction (Table 5) and zero-shot synthesis (Table 6). The only load-bearing step that reduces to its own inputs is the WER column of Table 5 combined with the Abstract's 'preserves expressive content' claim: the codec is trained with an ASR transcription objective, so low WER on decoded audio is the trained behavior, as evidenced by QDAC beating the reported ground-truth WER. This is genuine but partial: PESQ, STOI, SI-SDR, STFT/Mel, and the MOS/speaker-similarity results are not trained objectives and independently support the multi-codebook direction. The asserted semantic/acoustic disentanglement is an unsupported premise (no ablation, no mutual-information test), which is a correctness risk rather than circularity. The Section 3.2 claim that the 16-codebook model 'substantially surpasses the single-codebook baselines on all metrics' is contradicted by Table 6, which reports only QTTS-8 and shows it worse than CosyVoice2 on WER, PGC-hard, and speaker similarity; this is an evidence/verifiability defect, not circularity. No self-citations are load-bearing; cited foundations (DAC, Encodec, Moshi, CosyVoice, SpeechTokenizer) are external prior work. Hence: one fitted-input-called-prediction step, score 5.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper's central claims rest on design choices such as codebook count, delay step, frame rate, and codebook size, and on untested assumptions about semantic and acoustic separation and metric validity. No independent evidence is provided for the disentanglement assumption beyond the ASR loss itself.

free parameters (4)
  • Number of RVQ codebooks K = 8 and 16 used; K=8 in the TTS table
    Core to the claim that more codebooks improve quality; selected per experiment without an optimality criterion or theoretical derivation. No 16-codebook TTS row is shown.
  • Codebook size and embedding dimension = 2048 entries, 1024-dimensional embeddings
    Inherited from DAC and not ablated, yet it determines bitrate and capacity of the representation.
  • Token frame rate = 25 Hz default, 50 Hz in some codec tests
    Trades reconstruction quality against sequence length; chosen by hand and not systematically optimized.
  • Multihead delay step k = not stated
    The delay offset in Multihead-Delay is a fixed number of steps k, but the paper never gives its value or an ablation, despite it being the key hyperparameter of that architecture.
assumptions (5)
  • domain assumption Cross-entropy loss through the first codebook forces semantic content into C1 and leaves C2...CN free to encode residual acoustic detail.
    Invoked in Section 2.1.2; no ablation or mutual-information analysis demonstrates that semantic and acoustic information are actually separated across codebooks.
  • domain assumption Autoregressive modeling over time and codebook depth captures all inter-codebook dependencies needed for high-fidelity synthesis.
    Used to justify Hierarchical Parallel in Section 2.3; no convergence or scaling analysis is provided.
  • domain assumption Fixed-step delay in Multihead-Delay preserves sufficient local context without long-range cross-codebook information.
    Assumed in Section 2.2; the delay value and its effect are not empirically isolated.
  • domain assumption Objective metrics PESQ, STOI, SI-SDR, STFT distance, Mel distance, WER, speaker similarity, and MOS are valid and comparable proxies for TTS quality.
    Used throughout Section 3; no listener details, confidence intervals, or significance tests are reported.
  • standard math RVQ residual cascade with argmin over prototypes converges to a progressively finer approximation of the encoder latent.
    Stated in Section 2.1.1; standard property of RVQ used in SoundStream, EnCodec, and DAC.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantize More, Lose Less: Autoregressive Generation from Residually Quantized Speech Representations." pith.science (2026). https://pith.science/paper/2Z4JPQFB

@misc{pith2026250712197,
  author       = {Pith},
  title        = {Pith review of: Quantize More, Lose Less: Autoregressive Generation from Residually Quantized Speech Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2Z4JPQFB}},
  note         = {Machine review of arXiv:2507.12197}
}
read the original abstract

Text-to-speech (TTS) synthesis has seen renewed progress under the discrete modeling paradigm. Existing autoregressive approaches often rely on single-codebook representations, which suffer from significant information loss. Even with post-hoc refinement techniques such as flow matching, these methods fail to recover fine-grained details (e.g., prosodic nuances, speaker-specific timbres), especially in challenging scenarios like singing voice or music synthesis. We propose QTTS, a novel TTS framework built upon our new audio codec, QDAC. The core innovation of QDAC lies in its end-to-end training of an ASR-based auto-regressive network with a GAN, which achieves superior semantic feature disentanglement for scalable, near-lossless compression. QTTS models these discrete codes using two innovative strategies: the Hierarchical Parallel architecture, which uses a dual-AR structure to model inter-codebook dependencies for higher-quality synthesis, and the Delay Multihead approach, which employs parallelized prediction with a fixed delay to accelerate inference speed. Our experiments demonstrate that the proposed framework achieves higher synthesis quality and better preserves expressive content compared to baseline. This suggests that scaling up compression via multi-codebook modeling is a promising direction for high-fidelity, general-purpose speech and audio generation.

Figures

Figures reproduced from arXiv: 2507.12197 by the authors.

Figure 1
Figure 1. Architecture of the audio tokenizer model. It adopts a GAN-based framework consisting of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison of two multi-codebook prediction strategies for TTS: (Left) The Parallel [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 1 canonical work pages

  1. [1]

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

    Philip Anastassiou, Jiawei Chen, Jitong Chen, Yuanzhe Chen, Zhuo Chen, Ziyi Chen, Jian Cong, Lelai Deng, Chuang Ding, Lu Gao, et al. Seed-tts: A family of high-quality versatile speech generation models. arXiv preprint arXiv:2406.02430,

    arXiv

  2. [5]

    High fidelity neural audio compression

    Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High fidelity neural audio compression. arXiv preprint arXiv:2210.13438,

    arXiv

  3. [6]

    Moshi: a speech-text foundation model for real-time dialogue

    Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave, and Neil Zeghidour. Moshi: a speech-text foundation model for real-time dialogue. arXiv preprint arXiv:2410.00037,

    arXiv

  4. [7]

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

    Zhihao Du, Qian Chen, Shiliang Zhang, Kai Hu, Heng Lu, Yexin Yang, Hangrui Hu, Siqi Zheng, Yue Gu, Ziyang Ma, et al. Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens. arXiv preprint arXiv:2407.05407, 2024a. Zhihao Du, Yuxuan Wang, Qian Chen, Xian Shi, Xiang Lv, Tianyu Zhao, Zhifu Gao, Yexin Yang, ...

    arXiv

  5. [8]

    Jaehyeon Kim, Jungil Kong, and Juhee Son

    URL https://arxiv.org/abs/2302.03540. Jaehyeon Kim, Jungil Kong, and Juhee Son. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In International Conference on Machine Learning , pages 5530–5540. PMLR,

    arXiv

  6. [10]

    Base tts: Lessons from building a billion-parameter text-to-speech model on 100k hours of data

    Mateusz Łajszczak, Guillermo Cámbara, Yang Li, Fatih Beyhan, Arent Van Korlaar, Fan Yang, Arnaud Joly, Álvaro Martín-Cortinas, Ammar Abbas, Adam Michalski, et al. Base tts: Lessons from building a billion-parameter text-to-speech model on 100k hours of data. arXiv preprint arXiv:2402.08093,

    arXiv

  7. [11]

    Julian D Parker, Anton Smirnov, Jordi Pons, CJ Carr, Zack Zukowski, Zach Evans, and Xubo Liu

    URL https://arxiv.org/abs/2306.07691. Julian D Parker, Anton Smirnov, Jordi Pons, CJ Carr, Zack Zukowski, Zach Evans, and Xubo Liu. Scaling transformers for low-bitrate high-quality speech coding,

    arXiv

  8. [12]

    org/abs/2411.19842

    URL https://arxiv. org/abs/2411.19842. Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech,

    arXiv

Show all 18 references
  1. [13]

    Xinsheng Wang, Mingqi Jiang, Ziyang Ma, Ziyu Zhang, Songxiang Liu, Linqin Li, Zheng Liang, Qixi Zheng, Rui Wang, Xiaoqin Feng, et al

    URL https://arxiv.org/abs/2006.04558. Xinsheng Wang, Mingqi Jiang, Ziyang Ma, Ziyu Zhang, Songxiang Liu, Linqin Li, Zheng Liang, Qixi Zheng, Rui Wang, Xiaoqin Feng, et al. Spark-tts: An efficient llm-based text-to-speech model with single-stream decoupled speech tokens. arXiv ...

    2006 arXiv

  2. [15]

    Uniaudio: An audio foundation model toward universal audio generation

    Dongchao Yang, Jinchuan Tian, Xu Tan, Rongjie Huang, Songxiang Liu, Xuankai Chang, Jiatong Shi, Sheng Zhao, Jiang Bian, Xixin Wu, et al. Uniaudio: An audio foundation model toward universal audio generation. arXiv preprint arXiv:2310.00704,

    arXiv

  3. [16]

    Multi-band melgan: Faster waveform generation for high-quality text-to-speech

    Geng Yang, Shan Yang, Kai Liu, Peng Fang, Wei Chen, and Lei Xie. Multi-band melgan: Faster waveform generation for high-quality text-to-speech. In 2021 IEEE Spoken Language Technology Workshop (SLT), pages 492–498. IEEE,

    2021

  4. [17]

    Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi

    URL https://arxiv.org/abs/2502.04128. Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Sound- stream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495–507,

    arXiv

  5. [18]

    URL https: //arxiv.org/abs/2505.07916. 10 A Contributors Core Contributors: Yichen Han, Xiaoyang Hao, Keming Chen, Weibo Xiong, Jun He, Ruonan Zhang, Junjie Cao, Yue Liu∗1 Contributors: Bowen Li, Dongrui Zhang, Hui Xia, Huilei Fu, Kai Jia, Kaixuan Guo, Mingli Jin, Qingyun Meng...

    arXiv

  6. [2020]

    9 Kundan Kumar, Rithesh Kumar, Thibault De Boissiere, Lucas Gestin, Wei Zhen Teoh, Jose Sotelo, Alexandre De Brebisson, Yoshua Bengio, and Aaron C Courville

    URL https://arxiv.org/abs/2010.05646. 9 Kundan Kumar, Rithesh Kumar, Thibault De Boissiere, Lucas Gestin, Wei Zhen Teoh, Jose Sotelo, Alexandre De Brebisson, Yoshua Bengio, and Aaron C Courville. Melgan: Generative adversarial networks for conditional waveform synthesis. Advan...

    2010 arXiv

  7. [2022]

    doi: 10.1109/jstsp.2022.3188113

    ISSN 1941-0484. doi: 10.1109/jstsp.2022.3188113. URL http://dx.doi.org/10.1109/JSTSP.2022.3188113. Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao, Kai Yu, and Xie Chen. F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching. arXi...

    1941

  8. [2023]

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

    Edresson Casanova, Kelly Davis, Eren Gölge, Görkem Göknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua Meyer, Reuben Morais, Samuel Olayemi, et al. Xtts: a massively multilingual zero-shot text-to-speech model. arXiv preprint arXiv:2406.04904,

    arXiv

  9. [2024]

    Better speech synthesis through scaling

    James Betker. Better speech synthesis through scaling. arXiv preprint arXiv:2305.07243,

    arXiv

  10. [2025]

    Maskgct: Zero-shot text-to-speech with masked generative codec transformer

    Yuancheng Wang, Haoyue Zhan, Liwei Liu, Ruihong Zeng, Haotian Guo, Jiachen Zheng, Qiang Zhang, Xueyao Zhang, Shunsi Zhang, and Zhizheng Wu. Maskgct: Zero-shot text-to-speech with masked generative codec transformer. arXiv preprint arXiv:2409.00750,

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.