Pith. sign in

REVIEW 3 major objections 6 minor 4 cited by

The paper claims autoregressive TTS can skip discrete audio tokens and model a 2048×-compressed continuous latent space, reaching a 1.88% word error rate on LibriSpeech Subset-B while streaming the first audio frame in 96 ms.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

CLEAR is a zero-shot TTS model that autoregressively predicts compact continuous audio latents with a per-token rectified flow head, reaching 1.88% WER on LibriSpeech Subset-B with an RTF of 0.29 and a 96 ms streaming delay.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection CLEAR is a well-engineered continuous AR TTS system, but the SOTA claim rests on a plausible train/test contamination that the paper never addresses, and the WER below ground truth is a red flag. the 3 major comments →

arxiv 2508.19098 v1 pith:5LL7YWIG submitted 2025-08-26 eess.AS

CLEAR: Continuous Latent Autoregressive Modeling for High-quality and Low-latency Speech Synthesis

classification eess.AS
keywords continuous latent autoregressive modelzero-shot text-to-speechrectified flowvariational autoencoderstreaming speech synthesislow latencyflow matchingspeech language model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 claims that autoregressive text-to-speech does not need discrete audio tokens at all: modeling a heavily compressed continuous latent space directly is both more faithful and much faster. CLEAR's core move is a VAE that squeezes 16 kHz audio at 2048× compression into about 7.8 continuous vectors per second, which a unidirectional transformer predicts one at a time, and a lightweight per-token rectified-flow head that turns each hidden state into a distribution over the next latent. Because the flow head denoises each latent independently, the first audio chunk can be rendered in 96 ms, so synthesis can stream instead of waiting for the whole utterance. On LibriSpeech test-clean the authors report a 1.88% word error rate, UTMOS 4.22, and a real-time factor of 0.29, beating the compared two-stage and diffusion-based systems on robustness while using about 78 autoregressive steps per 10 seconds versus 400–620 for continuous-spectrogram baselines. If correct, the work turns the standard LM-plus-diffusion-refiner cascade into a single jointly trained pass.

Core claim

The central claim is that a single-stage autoregressive model can generate natural zero-shot speech directly in continuous latent space, provided the latents are compact and each per-step distribution is modeled by rectified flow rather than Gaussian or MSE assumptions. CLEAR pairs an enhanced wav-VAE (parameter-free shortcut connections let it reconstruct speech at 2048× downsampling, about 7.8 latents per second) with a unidirectional transformer that emits one hidden state per step and a six-block MLP rectified-flow head that denoises the next latent from that state. Training is single-stage — rectified-flow loss, cosine direction loss, logit-normal timestep sampling — with classifier-fre

What carries the argument

The load-bearing pairing is a high-compression waveform VAE with a per-token MLP rectified-flow head. The VAE's parameter-free shortcut connections (space-to-channel skips) let a convolutional encoder survive 2048× downsampling, yielding ~7.8 continuous latents per second. The flow head — six residual MLP blocks — takes one autoregressive hidden state h_k and a noise level and returns the vector field v(y_t^k|h_k,t), the denoising velocity for the next latent. Because each token's denoising depends only on its own hidden state, decoding begins before the full sequence exists; this unlocks streaming. The single-stage objective L_RF + L_D + L_stop combines rectified-flow loss, cosine direction

Load-bearing premise

The load-bearing premise is that scores quoted from other models — trained on different datasets and measured with different speech recognizers, with no confidence intervals on the small WER gaps — are directly comparable to CLEAR's, so a 0.5-point gap counts as a real quality edge rather than possible evaluation noise.

What would settle it

Re-train CLEAR-Large and its closest continuous competitor on identical data, decode both with the same recognizer, and report WER, SIM-o, and UTMOS with confidence intervals across several seeds: if the 0.51-point WER gap (1.88 vs 2.39) collapses to within noise, the state-of-the-art claim is unsupported. Separately, reproduce the streaming claim by measuring first-frame latency on the target hardware with chunking and overlap fades enabled, checking that the 96 ms figure holds under real-time load.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Two-stage designs — a codec LM followed by a separate diffusion refiner — become unnecessary: CLEAR trains one model end-to-end and still reports UTMOS 4.21–4.27, so the cascade adds complexity without a clear quality payoff in this comparison.
  • Streaming synthesis at 96 ms first-frame latency becomes practical because the MLP flow head denoises each latent as soon as its hidden state exists, and the causal VAE decoder renders chunk-by-chunk with overlap fades.
  • Autoregressive decoding drops to about 78 steps per 10 seconds of audio, driving the real-time factor to 0.18 (Base) and 0.29 (Large), so high-quality zero-shot TTS fits on consumer GPUs.
  • The compression ratio is a tunable dial: raising it from 768 to 4096 cuts WER from 15.4% to 4.90% at the cost of speaker similarity, so a 2048× ratio is a deliberate middle point.
  • Discrete audio codecs — with their bitrate limits, codebook training sensitivity, and lossy compression — become optional in the AR TTS stack; a continuous VAE plus per-token flow head covers the same ground.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same recipe — high-ratio continuous VAE plus per-token flow head — should transfer to other autoregressive audio generation tasks (music, sound effects, voice conversion) that currently pay the same discrete-codec latency and bitrate costs; nothing in the method is speech-specific.
  • Because the VAE emits ~7.8 frames per second while a mel spectrogram needs 100, the AR transformer sees a sequence roughly 13× shorter; the efficiency gain should grow with model scale since attention cost is sequence-length dependent.
  • The paper's own SIM-o gap (0.59 vs 0.65–0.67 for leading systems) names a concrete extension the authors flag in their limitations: injecting a speaker embedding into both the LM and the flow head should recover speaker similarity without sacrificing the compression benefits.
  • The 96 ms figure is measured on one GPU configuration; on-device latency will depend on the causal decoder's chunk rendering and the fade-in/fade-out overlap, so a hardware benchmark beyond the reported FFL is the natural next measurement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes CLEAR, a zero-shot TTS model that autoregressively predicts continuous audio latents rather than discrete codec tokens. Speech is encoded by a high-compression variational autoencoder (downsampling ratio 2048) into about 7.8 latents per second; an autoregressive Transformer produces per-token conditioning vectors, and a lightweight MLP rectified-flow head models the conditional distribution of the next continuous latent. The AR model and flow head are trained jointly in a single stage, with classifier-free guidance at inference. The authors report competitive or state-of-the-art objective scores on LibriSpeech test-clean subsets (e.g., 1.88% WER and 4.22 UTMOS for CLEAR-Large on Subset-B), an RTF of 0.18–0.29, and streaming synthesis with a 96 ms first-frame delay. Appendices include ablations on compression ratio, CFG/training details, subjective listening protocols, and a candid discussion of the model's lower objective speaker similarity.

Significance. If the reported evaluations are valid, CLEAR makes a useful contribution: it shows that continuous-latent autoregressive TTS can be trained in one stage, that a compact VAE latent sequence drastically reduces the number of AR decoding steps, and that a per-token MLP flow head enables streaming synthesis with low first-frame latency. The architecture is clearly described, the ablations (compression ratio, CFG strategy, model size) are informative, and the authors explicitly acknowledge the speaker-similarity limitation in Appendix A and Appendix D.2. The main value is the efficiency/latency argument rather than a fundamentally new modeling paradigm. However, the SOTA claim currently rests on cross-paper point estimates and on an evaluation dataset that may overlap the training corpus; both issues must be resolved before the headline numbers can be accepted.

major comments (3)
  1. [Section 4.1, Table 1] Train/test overlap is not addressed. CLEAR is trained on LibriHeavy, which is derived from LibriVox audiobooks, and evaluated on LibriSpeech test-clean / LibriSpeech-PC test-clean subsets, which are themselves LibriVox subsets. The paper reports no utterance-level or speaker-level de-duplication between training and evaluation. This matters concretely: CLEAR-Large's WER (1.88%) is lower than the ground-truth WER (2.47%) and the VAE-reconstruction WER (2.89%) on the same Subset-B. While synthetic speech can occasionally be more ASR-intelligible than original audio, a below-ground-truth WER combined with a known LibriVox derivation is exactly the signature of memorized test content. Please report the exact overlap (utterance IDs and speaker IDs) between LibriHeavy and the two evaluation subsets, and re-evaluate on a disjoint held-out LibriVox set if any overlap exists. Without this, the ce
  2. [Section 5.1, Tables 1 and 2] The headline comparisons rely on quoted baseline numbers that are not directly comparable. Baselines in Table 1 are trained on different corpora (Libri-60k, Emilia-100k, Multi-170k) and some scores are quoted from papers using different ASR/protocols; footnotes mix quoted and reproduced values. The WER differences are small (e.g., 1.88 vs. 2.39 vs. DiTAR on Subset-B), but no confidence intervals, bootstrap intervals, or significance tests are provided. The 0.51% absolute WER reduction claimed in Section 5.1(c) could be within evaluation noise. Please provide confidence intervals for the CLEAR numbers and for the strongest baselines, and run at least DiTAR, F5-TTS, and CosyVoice 2 under the same ASR and prompt protocol on the same subsets, or explicitly restrict the claims to 'competitive' rather than 'SOTA'.
  3. [Section 5.1, Table 2; Section 5.3] The RTF and latency comparisons lack the measurement conditions needed to support the 'best RTF' claim. Table 2 quotes RTF values from other papers (VALL-E R, CLaM-TTS, F5-TTS, DiTAR, MELLE) without specifying hardware, batch size, NFE, or optimization settings for those baselines, and the CLEAR RTF measurement is not described beyond a later mention of 'a 4096-GPU' in Section 5.3. Similarly, the 96 ms first-frame latency in Table 4 is reported without a precise measurement protocol (GPU model, chunk encoding/denoising schedule, overlap handling). Please state the hardware and measurement conditions for all systems, or avoid comparative RTF claims.
minor comments (6)
  1. [Section 5.1] The text says 'reduces the required autoregressive decoding steps to as low as 7.8', but Table 2 reports 78 average decoding steps for a 10-second segment. 7.8 is the number of latents per second; please correct this inconsistency.
  2. [Section 3.1, Figure 1] The token labels 'S' and 'T' are described as 'start of sequence' and 'turn of speech', but the input sequence in Section 3.5 is written as [S, x, T, y]. Clarify what 'T' means in the sequence notation.
  3. [Abstract / Introduction] Several typos and formatting issues: 'from a few seconds of a audio prompt', 'and and a causal VAE decoder', 'regrading Transformer as denosing backbone', and 'Closed-sourced' in Table 1. These should be cleaned up.
  4. [Section 4.1 / Table 1] The dataset naming is inconsistent: 'LibriSpeech(PC) test-clean', 'LibriSpeech-PC test-clean', and 'LibriSpeech test-clean' are used interchangeably. Please define the exact relationship and use one consistent name.
  5. [Appendix D.2, Table 8] The second speaker-similarity column is labeled only 'WavLM'; the text refers to WavLM-base-sv. Please make the column header explicit.
  6. [Abstract / Audio link] 'Audio samples are available at here' is an incomplete reference. No URL or availability statement for code/pretrained models is given, which limits reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the paper's WER/UTMOS/RTF results are external measurements, not quantities derived from its own fitted parameters or self-citations.

full rationale

CLEAR is an empirical systems paper: its central claims are objective/subjective measurements on LibriSpeech test-clean subsets (Table 1), efficiency measurements (Table 2), and streaming-latency measurements (Table 4). These are not derived from the model equations by construction. The autoregressive factorization in Eq. (1), the rectified-flow loss in Eq. (3), the auxiliary direction loss in Eq. (4), and the CFG blending in Eq. (6) are standard training objectives; none of them defines a target quantity in terms of itself. The VAE and flow-head hyperparameters (e.g., compression ratio D=2048, CFG scale 2.5) are selected via ablations and implementation choices, but they are not fitted parameters that are later relabeled as predictions. The paper contains no load-bearing self-citation chain: its architectural references (e.g., Deep Compression Autoencoder [5], rectified flow [34], oobleck blocks [15], snake activation [68]) are external prior work, not the present authors' own unverified claims. The skeptical concern about possible train/test overlap between LibriHeavy (LibriVox-derived) and LibriSpeech test-clean, and the striking fact that CLEAR-Large WER (1.88%) is below ground-truth WER (2.47%), is a data-contamination/correctness risk rather than a circularity: it does not make any reported quantity algebraically equal to its own input. Overall, the derivation chain is self-contained; no circular step can be exhibited with the required quote-and-reduction evidence.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central claim rests on a modest set of hand-chosen hyperparameters (compression ratio, guidance scale, NFE, loss weights, streaming chunk size) and on standard assumptions about rectified flow, the autoregressive factorization, and the validity of the evaluation metrics. No new physical or conceptual entities are introduced. The most fragile assumption is the comparability of quoted baselines, which directly supports the SOTA claim.

free parameters (7)
  • VAE downsampling ratio D = 2048
    Chosen from an ablation (Table 7) that trades WER versus SIM-o; directly sets the AR sequence length and hence RTF.
  • CFG guidance scale w = 2.5
    Inference hyperparameter in Eq. 6 balancing diversity and fidelity; no validation sweep reported.
  • Number of denoising function evaluations NFE = 10
    Sets the number of rectified flow steps per token, trading quality against latency.
  • Logit-normal timestep sampling parameters m, s = m=0, s=1
    Parameters of the timestep sampling distribution in Eq. 5, set without reported ablation.
  • CFG text-drop probability = 0.2
    Probability of zeroing text embeddings during training to enable classifier-free guidance.
  • VAE loss weights alpha, beta, gamma = alpha=1.0, beta=5.0, gamma=0.1
    Weights for feature-matching, adversarial, and KL losses in the VAE objective, chosen by hand.
  • Streaming chunk size and overlap = Omega=4, Psi=1
    Chunk size and overlap for streaming synthesis; Omega=4 gives 96 ms first-frame latency in Table 4.
axioms (5)
  • standard math Rectified flow can model arbitrary conditional densities p(y_k|h_k) via an ODE transport from noise to data.
    Invoked in Sec 3.2, Eq. (3), relying on the standard rectified flow framework.
  • domain assumption The autoregressive factorization p(y|x) = prod_k p(y_k|y<k, x) is valid for continuous speech latents.
    Sec 3.1 Eq. (1) assumes each latent frame is conditionally independent of all future frames given previous latents and text.
  • domain assumption Continuous latents at compression ratio 2048 retain enough information for high-quality TTS.
    Sec 3.4 and Appendix C.4.2 justify this via VAE reconstruction metrics, but it remains an assumed property of the representation.
  • domain assumption Objective metrics WER, SIM-o, and UTMOS are valid proxies for TTS quality.
    Sec 4.3 describes these metrics; the paper's quality claims depend on their validity, especially UTMOS as a naturalness predictor.
  • domain assumption Quoted baseline numbers from other papers are directly comparable to CLEAR's evaluations.
    Tables 1 and 2 compare against published numbers from models trained on different data, without re-running baselines in identical conditions.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CLEAR: Continuous Latent Autoregressive Modeling for High-quality and Low-latency Speech Synthesis." pith.science (2026). https://pith.science/paper/5LL7YWIG

@misc{pith2026250819098,
  author       = {Pith},
  title        = {Pith review of: CLEAR: Continuous Latent Autoregressive Modeling for High-quality and Low-latency Speech Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5LL7YWIG}},
  note         = {Machine review of arXiv:2508.19098}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Autoregressive (AR) language models have emerged as powerful solutions for zero-shot text-to-speech (TTS) synthesis, capable of generating natural speech from a few seconds of audio prompts. However, conventional AR-based TTS systems relying on discrete audio tokens face the challenge of lossy compression during tokenization, requiring longer discrete token sequences to capture the same information as continuous ones, which adds inference latency and complicates AR modeling. To address this challenge, this paper proposes the Continuous Latent Autoregressive model (CLEAR), a unified zero-shot TTS framework that directly models continuous audio representations. More specifically, CLEAR introduces an enhanced variational autoencoder with shortcut connections, which achieves a high compression ratio to map waveforms into compact continuous latents. A lightweight MLP-based rectified flow head that operates independently for each hidden state is presented to model the continuous latent probability distribution, and trained jointly with the AR model within a single-stage framework. Experiments show that the proposed zero-shot CLEAR TTS can synthesize high-quality speech with low latency. Compared to state-of-the-art (SOTA) TTS models, CLEAR delivers competitive performance in robustness, speaker similarity and naturalness, while offering a lower real-time factor (RTF). In particular, CLEAR achieves SOTA results on the LibriSpeech test-clean dataset, with a word error rate of 1.88\% and an RTF of 0.29. Moreover, CLEAR facilitates streaming speech synthesis with a first-frame delay of 96ms, while maintaining high-quality speech synthesis.

Figures

Figures reproduced from arXiv: 2508.19098 by Chun Yat Wu, Guinan Li, Jiajun Deng, Qiuqiang Kong, Simon Lui.

Figure 1
Figure 1. Figure 1: Overview of the CLEAR architecture (left) and the enhanced wav-VAE architecture (right). [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The parameter-free shortcut connection used in waveform VAEs. For space-to-channel [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Model convergence of our continuous VAE (KL divergence regularization) and discrete [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: CLEAR-Base model performance (WER(%)/SIM-o) versus training steps. [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The user interface for subjective evaluation. [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The questions of each samples pair for subjective evaluation. When calculating the MOS [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 4 Pith papers

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

  1. SemaVoice: Semantic-Aware Continuous Autoregressive Speech Synthesis

    eess.AS 2026-05 unverdicted novelty 6.0

    SemaVoice adds SFM-guided alignment to refine continuous speech representations in autoregressive TTS, reporting 1.71% English WER on Seed-TTS and competitiveness with open-source SOTA.

  2. VoxCPM2 Technical Report

    cs.SD 2026-06 unverdicted novelty 5.0

    VoxCPM2 scales hierarchical continuous-latent speech modeling to 2B parameters and over 2M hours of multilingual data, unifying voice cloning, style control, and continuation in one backbone with open release.

  3. On the Distillation Loss Functions of Speech VAE for Unified Reconstruction, Understanding, and Generation

    cs.SD 2026-04 unverdicted novelty 5.0

    Joint-marginal alignment plus adaptive weighting in speech VAE distillation yields the best combined performance on reconstruction, understanding, and generation tasks.

  4. On the Distillation Loss Functions of Speech VAE for Unified Reconstruction, Understanding, and Generation

    cs.SD 2026-04 unverdicted novelty 4.0

    Joint-marginal distillation with adaptive weighting is reported as the best overall alignment loss for speech VAEs across reconstruction, understanding, and generation.

Reference graph

Works this paper leans on

69 extracted references · 34 canonical work pages · cited by 3 Pith papers

  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, 2024

  2. [2]

    SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing, May 2022

    Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, and Furu Wei. SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing, May 2022

  3. [3]

    Rethinking lossy compression: The rate-distortion-perception tradeoff

    Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In International Conference on Machine Learning, pages 675–685. PMLR, 2019

  4. [4]

    Audiolm: a language modeling approach to audio generation

    Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, et al. Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing, 31:2523–2533, 2023

  5. [5]

    Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models, April 2025

    Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models, April 2025

  6. [6]

    Wavlm: Large-scale self-supervised pre- training for full stack speech processing

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, et al. Wavlm: Large-scale self-supervised pre- training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 16(6):1505–1518, 2022

  7. [7]

    Vall-e 2: Neural codec language models are human parity zero-shot text to speech synthesizers, 2024

    Sanyuan Chen, Shujie Liu, Long Zhou, Yanqing Liu, Xu Tan, Jinyu Li, Sheng Zhao, Yao Qian, and Furu Wei. Vall-e 2: Neural codec language models are human parity zero-shot text to speech synthesizers, 2024. URL https://arxiv.org/abs/2406.05370

  8. [8]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching, 2024

    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, 2024. URL https://arxiv.org/abs/2410.06885

  9. [9]

    High Fidelity Neural Audio Compression, October 2022

    Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High Fidelity Neural Audio Compression, October 2022

  10. [10]

    Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens, 2024

    Zhihao Du, Qian Chen, Shiliang Zhang, Kai Hu, Heng Lu, Yexin Yang, Hangrui Hu, Siqi Zheng, Yue Gu, Ziyang Ma, Zhifu Gao, and Zhijie Yan. Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens, 2024. URL https: //arxiv.org/abs/2407.05407

  11. [11]

    Cosyvoice 2: Scalable streaming speech synthesis with large language models, 2024

    Zhihao Du, Yuxuan Wang, Qian Chen, Xian Shi, Xiang Lv, Tianyu Zhao, Zhifu Gao, Yexin Yang, Changfeng Gao, Hui Wang, Fan Yu, Huadai Liu, Zhengyan Sheng, Yue Gu, Chong Deng, Wen Wang, Shiliang Zhang, Zhijie Yan, and Jingren Zhou. Cosyvoice 2: Scalable streaming speech synthesis with large language models, 2024. URL https://arxiv.org/abs/2412.10117

  12. [12]

    E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts, 2024

    Sefik Emre Eskimez, Xiaofei Wang, Manthan Thakker, Canrun Li, Chung-Hsien Tsai, Zhen Xiao, Hemin Yang, Zirun Zhu, Min Tang, Xu Tan, Yanqing Liu, Sheng Zhao, and Naoyuki Kanda. E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts, 2024. URL https: //arxiv.org/abs/2406.18009

  13. [13]

    Taming Transformers for High-Resolution Image Synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming Transformers for High-Resolution Image Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873–12883, 2021. 10

  14. [14]

    Scaling rectified flow transformers for high-resolution image synthesis, 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. URL https://arxiv.org/abs/ 2403.03206

  15. [15]

    Hawley, and Jordi Pons

    Zach Evans, CJ Carr, Josiah Taylor, Scott H. Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion, 2024. URL https://arxiv.org/abs/2402.04825

  16. [16]

    E3 TTS: Easy End-to-End Diffusion-Based Text To Speech

    Yuan Gao, Nobuyuki Morioka, Yu Zhang, and Nanxin Chen. E3 TTS: Easy End-to-End Diffusion-Based Text To Speech. In ASRU, pages 1–8, February 2023. doi: 10.1109/ ASRU57964.2023.10389766

  17. [17]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks, June 2014

  18. [18]

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

    Hao-Han Guo, Yao Hu, Kun Liu, Fei-Yu Shen, Xu Tang, Yi-Chen Wu, Feng-Long Xie, Kun Xie, and Kai-Tuo Xu. Fireredtts: A foundation text-to-speech framework for industry-level generative speech applications. arXiv preprint arXiv:2409.03283, 2024

  19. [19]

    Vall-e r: Robust and efficient zero-shot text-to-speech synthesis via monotonic alignment

    Bing Han, Long Zhou, Shujie Liu, Sanyuan Chen, Lingwei Meng, Yanming Qian, Yanqing Liu, Sheng Zhao, Jinyu Li, and Furu Wei. Vall-e r: Robust and efficient zero-shot text-to-speech synthesis via monotonic alignment. arXiv preprint arXiv:2406.07855, 2024

  20. [20]

    Classifier-free diffusion guidance, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. URL https:// arxiv.org/abs/2207.12598

  21. [21]

    Straightening out the straight- through estimator: Overcoming optimization challenges in vector quantized networks

    Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight- through estimator: Overcoming optimization challenges in vector quantized networks. In International Conference on Machine Learning, pages 14096–14113. PMLR, 2023

  22. [22]

    Ditar: Diffusion transformer au- toregressive modeling for speech generation, 2025

    Dongya Jia, Zhuo Chen, Jiawei Chen, Chenpeng Du, Jian Wu, Jian Cong, Xiaobin Zhuang, Chumin Li, Zhen Wei, Yuping Wang, and Yuxuan Wang. Ditar: Diffusion transformer au- toregressive modeling for speech generation, 2025. URL https://arxiv.org/abs/2502. 03930

  23. [23]

    Mega-TTS 2: Boosting Prompting Mechanisms for Zero-Shot Speech Synthesis, April 2024

    Ziyue Jiang, Jinglin Liu, Yi Ren, Jinzheng He, Zhenhui Ye, Shengpeng Ji, Qian Yang, Chen Zhang, Pengfei Wei, Chunfeng Wang, Xiang Yin, Zejun Ma, and Zhou Zhao. Mega-TTS 2: Boosting Prompting Mechanisms for Zero-Shot Speech Synthesis, April 2024

  24. [24]

    NaturalSpeech 3: Zero-Shot Speech Synthesis with Factorized Codec and Diffusion Models, April 2024

    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. NaturalSpeech 3: Zero-Shot Speech Synthesis with Factorized Codec and Diffusion Models, April 2024

  25. [25]

    Libriheavy: a 50,000 hours asr corpus with punctuation casing and context,

    Wei Kang, Xiaoyu Yang, Zengwei Yao, Fangjun Kuang, Yifan Yang, Liyong Guo, Long Lin, and Daniel Povey. Libriheavy: a 50,000 hours asr corpus with punctuation casing and context,

  26. [26]

    Analyzing and Improving the Training Dynamics of Diffusion Models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and Improving the Training Dynamics of Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174–24184, 2024

  27. [27]

    Clam-tts: Improving neural codec language model for zero-shot text-to-speech

    Jaehyeon Kim, Keon Lee, Seungjun Chung, and Jaewoong Cho. Clam-tts: Improving neural codec language model for zero-shot text-to-speech. arXiv preprint arXiv:2404.02781, 2024

  28. [28]

    High-Fidelity Audio Compression with Improved RVQGAN

    Rithesh Kumar, Prem Seetharaman, Alejandro Luebs, Ishaan Kumar, and Kundan Kumar. High-Fidelity Audio Compression with Improved RVQGAN. Advances in Neural Information Processing Systems, 36:27980–27993, December 2023

  29. [29]

    BASE TTS: Lessons from building a billion- parameter Text-to-Speech model on 100K hours of data, February 2024

    Mateusz Łajszczak, Guillermo Cámbara, Yang Li, Fatih Beyhan, Arent van Korlaar, Fan Yang, Arnaud Joly, Álvaro Martín-Cortinas, Ammar Abbas, Adam Michalski, Alexis Moinet, Sri Karlapati, Ewa Muszy´nska, Haohan Guo, Bartosz Putrycz, Soledad López Gambino, Kayeon Yoo, Elena Sokolova, and Thomas Drugman. BASE TTS: Lessons from building a billion- parameter Te...

  30. [30]

    V oicebox: Text-guided multilin- gual universal speech generation at scale

    Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, et al. V oicebox: Text-guided multilin- gual universal speech generation at scale. Advances in neural information processing systems, 36:14005–14034, 2023

  31. [31]

    REPA-E: Unlocking V AE for End-to-End Tuning with Latent Diffusion Transformers, April 2025

    Xingjian Leng, Jaskirat Singh, Yunzhong Hou, Zhenchang Xing, Saining Xie, and Liang Zheng. REPA-E: Unlocking V AE for End-to-End Tuning with Latent Diffusion Transformers, April 2025

  32. [32]

    Neural Speech Synthesis with Transformer Network

    Naihan Li, Shujie Liu, Yanqing Liu, Sheng Zhao, and Ming Liu. Neural Speech Synthesis with Transformer Network. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01): 6706–6713, July 2019. ISSN 2374-3468. doi: 10.1609/aaai.v33i01.33016706

  33. [33]

    Autoregressive image generation without vector quantization, 2024

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization, 2024. URL https://arxiv.org/abs/2406.11838

  34. [34]

    Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. URL https://arxiv.org/abs/2209.03003

  35. [35]

    Autoregressive diffusion transformer for text-to-speech synthesis, 2024

    Zhijun Liu, Shuai Wang, Sho Inoue, Qibing Bai, and Haizhou Li. Autoregressive diffusion transformer for text-to-speech synthesis, 2024. URL https://arxiv.org/abs/2406.05551

  36. [36]

    LibriSpeech-PC: Benchmark for Evaluation of Punctuation and Capitalization Capabilities of End-to-End ASR Models

    Aleksandr Meister, Matvei Novikov, Nikolay Karpov, Evelina Bakhturina, Vitaly Lavrukhin, and Boris Ginsburg. LibriSpeech-PC: Benchmark for Evaluation of Punctuation and Capitalization Capabilities of End-to-End ASR Models. In 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 1–7, February 2023

  37. [37]

    Autoregressive speech synthesis without vector quantization, 2024

    Lingwei Meng, Long Zhou, Shujie Liu, Sanyuan Chen, Bing Han, Shujie Hu, Yanqing Liu, Jinyu Li, Sheng Zhao, Xixin Wu, Helen Meng, and Furu Wei. Autoregressive speech synthesis without vector quantization, 2024. URL https://arxiv.org/abs/2407.08551

  38. [38]

    Finite Scalar Quantization: VQ-V AE Made Simple, October 2023

    Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite Scalar Quantization: VQ-V AE Made Simple, October 2023

  39. [39]

    Hall-e: hierarchical neural codec language model for minute-long zero-shot text-to-speech synthesis

    Yuto Nishimura, Takumi Hirose, Masanari Ohi, Hideki Nakayama, and Nakamasa Inoue. Hall-e: hierarchical neural codec language model for minute-long zero-shot text-to-speech synthesis. arXiv preprint arXiv:2410.04380, 2024

  40. [40]

    Librispeech: An ASR corpus based on public domain audio books

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: An ASR corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206–5210, April 2015

  41. [41]

    Robust speech recognition via large-scale weak supervision

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492–28518. PMLR, 2023

  42. [42]

    Revisiting over-smoothness in text to speech

    Yi Ren, Xu Tan, Tao Qin, Zhou Zhao, and Tie-Yan Liu. Revisiting over-smoothness in text to speech. In ACL, pages 8197–8213, 2022

  43. [43]

    Utmos: Utokyo-sarulab system for voicemos challenge 2022

    Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari. Utmos: Utokyo-sarulab system for voicemos challenge 2022. arXiv preprint arXiv:2204.02152, 2022

  44. [44]

    AutoClip: Adaptive Gradient Clipping for Source Separation Networks, July 2020

    Prem Seetharaman, Gordon Wichern, Bryan Pardo, and Jonathan Le Roux. AutoClip: Adaptive Gradient Clipping for Source Separation Networks, July 2020

  45. [45]

    Glu variants improve transformer, 2020

    Noam Shazeer. Glu variants improve transformer, 2020. URL https://arxiv.org/abs/ 2002.05202

  46. [46]

    NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers, May 2023

    Kai Shen, Zeqian Ju, Xu Tan, Yanqing Liu, Yichong Leng, Lei He, Tao Qin, Sheng Zhao, and Jiang Bian. NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers, May 2023. 12

  47. [47]

    Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network

    Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874–1883, 2016

  48. [48]

    Ella-v: Stable neural codec language modeling with alignment-guided sequence reordering

    Yakun Song, Zhuo Chen, Xiaofei Wang, Ziyang Ma, and Xie Chen. Ella-v: Stable neural codec language modeling with alignment-guided sequence reordering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 25174–25182, 2025

  49. [49]

    Steinmetz, Jordi Pons, Santiago Pascual, and Joan Serrà

    Christian J. Steinmetz, Jordi Pons, Santiago Pascual, and Joan Serrà. Automatic Multitrack Mixing With A Differentiable Mixing Console Of Neural Audio Effects. In ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 71–75, June 2021

  50. [50]

    Roformer: Enhanced transformer with rotary position embedding, 2023

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2023. URL https://arxiv.org/ abs/2104.09864

  51. [51]

    NaturalSpeech: End-to-End Text-to-Speech Synthesis With Human-Level Quality

    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. NaturalSpeech: End-to-End Text-to-Speech Synthesis With Human-Level Quality. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4234–4245, June 2024. ISSN 1939-3539. doi: 10.11...

  52. [52]

    Continuous Speech Synthesis using per-token Latent Diffusion, October 2024

    Arnon Turetzky, Nimrod Shabtay, Slava Shechtman, Hagai Aronowitz, David Haws, Ron Hoory, and Avihu Dekel. Continuous Speech Synthesis using per-token Latent Diffusion, October 2024

  53. [53]

    Neural Discrete Representation Learning

    Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural Discrete Representation Learning. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017

  54. [54]

    Neural codec language models are zero-shot text to speech synthesizers

    Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111, 2023

  55. [55]

    Felle: Autoregressive speech synthesis with token-wise coarse-to-fine flow matching, 2025

    Hui Wang, Shujie Liu, Lingwei Meng, Jinyu Li, Yifan Yang, Shiwan Zhao, Haiyang Sun, Yanqing Liu, Haoqin Sun, Jiaming Zhou, Yan Lu, and Yong Qin. Felle: Autoregressive speech synthesis with token-wise coarse-to-fine flow matching, 2025. URL https://arxiv.org/ abs/2502.11128

  56. [56]

    Spark-tts: An efficient llm-based text-to-speech model with single-stream decoupled speech tokens

    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 preprint arXiv:2503.01710, 2025

  57. [57]

    MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer, October 2024

    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, October 2024

  58. [58]

    Towards audio language modeling–an overview

    Haibin Wu, Xuanjun Chen, Yi-Cheng Lin, Kai-wei Chang, Ho-Lam Chung, Alexander H Liu, and Hung-yi Lee. Towards audio language modeling–an overview. arXiv preprint arXiv:2402.13236, 2024

  59. [59]

    Rall-e: Robust codec lan- guage modeling with chain-of-thought prompting for text-to-speech synthesis

    Detai Xin, Xu Tan, Kai Shen, Zeqian Ju, Dongchao Yang, Yuancheng Wang, Shinnosuke Takamichi, Hiroshi Saruwatari, Shujie Liu, Jinyu Li, et al. Rall-e: Robust codec lan- guage modeling with chain-of-thought prompting for text-to-speech synthesis. arXiv preprint arXiv:2404.03204, 2024

  60. [60]

    On layer normalization in the transformer architecture, 2020

    Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tie-Yan Liu. On layer normalization in the transformer architecture, 2020. URL https://arxiv.org/abs/2002.04745

  61. [61]

    Reconstruction vs

    Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. Generation: Taming Opti- mization Dilemma in Latent Diffusion Models, March 2025. 13

  62. [62]

    Llasa: Scaling train-time and inference-time compute for llama-based speech synthesis

    Zhen Ye, Xinfa Zhu, Chi-Min Chan, Xinsheng Wang, Xu Tan, Jiahe Lei, Yi Peng, Haohe Liu, Yizhu Jin, Zheqi DAI, et al. Llasa: Scaling train-time and inference-time compute for llama-based speech synthesis. arXiv preprint arXiv:2502.04128, 2025

  63. [63]

    Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think, February 2025

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think, February 2025

  64. [64]

    Soundstream: An end-to-end neural audio codec

    Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495–507, 2021

  65. [65]

    Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu

    Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J. Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu. LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech, April 2019

  66. [66]

    Root mean square layer normalization, 2019

    Biao Zhang and Rico Sennrich. Root mean square layer normalization, 2019. URL https: //arxiv.org/abs/1910.07467

  67. [67]

    Speak foreign languages with your own voice: Cross-lingual neural codec language modeling

    Ziqiang Zhang, Long Zhou, Chengyi Wang, Sanyuan Chen, Yu Wu, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al. Speak foreign languages with your own voice: Cross-lingual neural codec language modeling. arXiv preprint arXiv:2303.03926, 2023

  68. [68]

    ,𝑁𝑆𝐵,𝐶!,𝑁 + UpsamplingBlock Channel toSpace ChannelDuplicating𝐵,𝐶

    Liu Ziyin, Tilman Hartwig, and Masahito Ueda. Neural Networks Fail to Learn Periodic Functions and How to Fix It. InAdvances in Neural Information Processing Systems, volume 33, pages 1583–1594. Curran Associates, Inc., 2020. 14 A Limitations Although CLEAR achieves competitive results with an efficient architecture, we identify several areas for potentia...

  69. [2024]

    URL https://arxiv.org/abs/2309.08105

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.