Pith. sign in

REVIEW 4 major objections 5 minor 23 references

Sparse temporal embedding replaces rigid alignment with a single anchor per phone, yielding a TTS that is both faster and less error-prone than larger alternatives.

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 →

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

T0 review · deepseek-v4-flash

2026-08-01 11:28 UTC pith:3RQVBGFS

load-bearing objection A real efficiency win with a novel conditioning trick, but the robustness claim needs a duration-predictor sanity check before I'd trust it. the 4 major comments →

arxiv 2607.19859 v1 pith:3RQVBGFS submitted 2026-07-22 cs.SD

StellarTTS: Sparse Temporal Embedding for Low-Latency and Robust Speech Synthesis

Kaicheng Luo , Xuefei Gong , Yutao Sun , Jinling He , Yujie Hou , Xiaoyang Xing , Huiyan Li , Bing Han
show 1 more author
Yanmin Qian
This is my paper
classification cs.SD
keywords zero-shot TTSnon-autoregressive TTSsparse temporal embeddingsemantic-aware codecmasked generative transformerlow latencyrobustnessmobile deployment
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 paper sets out to dissolve the long-standing trade-off between robustness, latency, and prosodic naturalness in text-to-speech. It claims that a non-autoregressive masked-transformer TTS can be both fast and accurate if, instead of enforcing frame-by-frame phone alignment, the model is given only a sparse temporal embedding: each phone's identity appears at a single central time step and all other positions carry a padding token. Combined with a semantic-aware codec that folds semantic and acoustic token prediction into one stage, an 83M-parameter system reaches a real-time factor of 0.08 and lowers word error rates below larger baselines on both ordinary and deliberately difficult Chinese test sentences. A sympathetic reader would care because this suggests high-quality, low-latency speech synthesis is achievable at a size and speed compatible with mobile devices.

Core claim

The discovery is that the position of the anchor in a sparse temporal embedding carries the linguistic structure: placing each phone's embedding exactly at the center of its duration, rather than across the whole span or at a random offset, is what keeps the model robust when all other temporal positions are non-informative. With ground-truth durations during training and a predictor that places the anchor at half the predicted length during inference, the model achieves a WER of 1.44 on Seed-TTS test-zh and 7.47 on the harder test-hard, compared to 2.34/9.15 for a random anchor and 4.10/18.12 with no temporal embedding at all. The same system attains RTF 0.08 on a single A100, or about 4-9x

What carries the argument

Two mechanisms carry the argument. The first is the sparse temporal embedding: for each phone i with duration d_i, the phone embedding Embed(p_i) is placed only at position floor(d_i/2), and all other positions in that duration get a distinct padding embedding Embed(p_pad). This creates a sequence of isolated 'anchors' that preserve phone identities and their relative timing while leaving the model free to generate rich transitional prosody between anchors. An auxiliary duration predictor, trained with mean-squared error and gradient truncation so it does not back-propagate into the phone embeddings, estimates d_i at inference and so determines where the anchors go. The second is the semanti

Load-bearing premise

The load-bearing premise is that the duration predictor will place each anchor close enough to the true center of its phone at inference time; if predicted durations are systematically off, anchors may shift into neighboring phones and cascade into pronunciation errors.

What would settle it

Use the training-time ground-truth durations at inference instead of the duration predictor's estimates and compare WER on the Seed-TTS sets; alternately, deliberately shift every anchor by one or two frames. If WER rises sharply under either perturbation, the robustness claim depends on an accurate center placement rather than on the anchor mechanism itself.

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

If this is right

  • On the Seed-TTS test-zh and test-hard benchmarks, StellarTTS reports the lowest WER among compared systems (1.44 and 7.47), and its CMOS is competitive with or above the best baselines.
  • The model's RTF of 0.08 on a single A100 represents a 4-9x speed-up over MaskGCT and F5-TTS, making real-time mobile deployment realistic.
  • Phoneme-level duration control (scaling predicted durations by 0.7 to 1.3) keeps WER below the best MaskGCT result at every multiplier, supporting human refinement of speed.
  • Ablations show that removing the temporal embedding or choosing a random anchor dramatically worsens WER, pinpointing the central anchor as the source of robustness.
  • The single-stage semantic-aware codec eliminates the separate semantic-token prediction stage used by earlier masked-generative TTS pipelines.

Where Pith is reading between the lines

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

  • The anchor-at-half-duration trick is a form of sparse conditioning that could be transferred to other sequence-to-sequence generation tasks where events have variable durations, such as gesture generation or music synthesis, if the center position carries analogous temporal information.
  • A natural stress test would be to measure WER under deliberate perturbations of the duration predictor's outputs; the paper does not report prediction error, so this residual is the main unknown failure mode introduced by moving from full alignment to sparse anchoring.
  • The small speaker-similarity gap (SIM-o 0.712 vs MaskGCT's 0.774) is attributed by the authors to the codec's semantic distillation; future work might try distilling to a later RVQ channel or adding a speaker-aware regularization to recover SIM-o without rebuilding multi-stage decoding.
  • Setting inference steps to [8,4,1,1,1,1] across RVQ layers suggests the approach's latency scales with number of refinement steps on channel 0; a testable extension is to see how few steps on channel 0 still preserve the WER advantage.

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

4 major / 5 minor

Summary. The paper proposes StellarTTS, a non-autoregressive zero-shot TTS system whose key design is a "sparse temporal embedding": for each phone, a single phone embedding is placed at the frame floor(d_i/2) while all other frames receive a padding embedding. This sparse anchor is meant to replace the rigid length-regulator alignment of classic NAR TTS and the alignment-free but less robust conditioning of masked generative models. The system also introduces a semantic-aware RVQ codec with one semantic channel distilled from Wav2vec-Bert and five acoustic channels, enabling single-stage decoding by an 83M-parameter masked generative transformer. The paper reports a real-time factor of 0.08 on an A100, WER of 1.44/7.47 on Seed-TTS test-zh/test-hard, and competitive CMOS/SMOS, and claims superior robustness and latency over AR/NAR baselines.

Significance. If the reported results hold, the sparse temporal embedding is a genuinely useful and conceptually clean design: it preserves phone-level controllability and linguistic robustness without imposing frame-wise deterministic prosody. The objective comparison against strong baselines (MaskGCT, F5-TTS, CosyVoice, FireRedTTS), the use of a hard test set, and the duration-control experiment in §IV-D are notable strengths, and the demo page is a useful asset. However, the current evidence is not yet at the level claimed: duration-predictor accuracy is load-bearing but unreported, comparative claims rest on point estimates without uncertainty, and the mobile-deployment claim is not supported by any measurement on mobile hardware. The significance would be materially higher if these gaps were closed.

major comments (4)
  1. [III-B, IV-D, Eqs. (3)-(4)] The sparse temporal embedding places the only phone-identity signal for each interval at position floor(d_i/2). At inference, predicted durations from the Section III-B predictor determine both the target sequence length and all anchor positions. The paper reports no duration-prediction error (MAE/RMSE), no oracle-vs-predicted duration comparison, and no analysis of how anchor misplacement degrades WER. The uniform duration scaling experiment in §IV-D preserves the relative anchor structure and is not equivalent to per-phone prediction errors. Because the central robustness claim is attributed to sparse temporal embedding, please report duration prediction accuracy and WER with ground-truth durations vs predicted durations.
  2. [Table I, IV-A, IV-B] All objective and subjective comparisons are point estimates with no error bars, confidence intervals, or significance tests. In Table I, the WER difference between StellarTTS (1.44), FireRedTTS (1.51), and F5-TTS (1.56) is small, and the CMOS differences in Table II are even smaller (e.g., +0.06 vs F5-TTS). Without uncertainty quantification or multiple runs, the claim of "stronger robustness" over state-of-the-art is not established. Please provide at least bootstrap confidence intervals or per-item standard errors.
  3. [Abstract, I, IV (Evaluation Metrics)] The abstract and introduction describe the system as "mobile-optimized" and cite the Qualcomm SM8650 as a deployment target, but RTF is computed on a single NVIDIA A100 GPU. No measurement on mobile hardware, no memory footprint, and no on-device latency are reported. Since low-latency mobile deployment is a central claim, please measure RTF and memory on the target mobile platform or clearly delimit the claim to server-side acceleration.
  4. [Table III, IV-C] The text in §IV-C describes a comparison between the "central strategy" and a "random strategy" (random anchor selection), but the ablation row in Table III is labeled "w/o central strategy." This is ambiguous: does the row correspond to random anchor selection, or to removal of the central anchor? The central-strategy ablation is one of the paper's core pieces of evidence, and the current label makes it impossible to tell which condition was actually run. Please correct the row label and clearly define each ablation condition.
minor comments (5)
  1. [III-A, Eq. (1)] The loss weights λ_rec, λ_codebook, λ_commit, and λ_distill are never specified, and the total codebook size per RVQ channel is not given. Please report these for reproducibility.
  2. [III-B] The duration predictor architecture is not described: input representation, hidden size, number of layers, and whether it is autoregressive over phones. Please add this detail.
  3. [III-C] The inference schedule [8,4,1,1,1,1] is said to total 16 steps; please clarify how these steps are distributed across the 6 RVQ channels and how the per-channel steps map to the cosine unmasking bound.
  4. [IV-B] The subjective evaluation uses 20 raters but no inter-rater agreement or confidence intervals are reported. Please state the rating procedure (e.g., 7-point scale, paired or absolute) and report at least standard errors.
  5. [Throughout] Minor typos and inconsistent capitalization: "comparative analyzes" should be "comparative analyses"; "0-th channel" should be "0th channel"; "Wav2vec-Bert" is sometimes written as "W2v-bert" in references.

Circularity Check

0 steps flagged

No significant circularity: the central claims are evaluated against external benchmarks and the proposed components are not defined in terms of the target results.

full rationale

The paper's main derived claims—WER 1.44/7.47, RTF 0.08, and CMOS/SMOS—are empirical results obtained on external Seed-TTS test-zh/test-hard sets using standard metrics (HuBERT-based ASR for WER, WavLM-TDCNN for SIM-o) and compared against independently published baselines. The semantic-aware codec is trained by KL distillation toward W2v-Bert features, while WER is calculated with HuBERT, so the evaluation metric is not the training target. The sparse temporal embedding of Eq. 3 is a construction: t_i determines sequence length and anchor position, so duration scaling in Section IV-D changes the input by design. This is a property of the method, not a fitted parameter renamed as a prediction. The duration predictor is an internal module trained with MSE (Eq. 4) on Emilia, not fitted to Seed-TTS test data, and no reported result is mathematically forced by the training objective. The absence of an oracle-vs-predicted duration error analysis is a validation gap, not circularity within the derivation chain. No load-bearing self-citation, uniqueness import, or ansatz-smuggling is present.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The paper contributes an empirical architecture; nothing is derived. Its main claims rest on hand-chosen losses, schedules, and pretrained feature extractors, with several key configuration values unreported.

free parameters (4)
  • codec loss weights λ_rec, λ_codebook, λ_commit, λ_distill = not reported
    Equation (1) defines a weighted sum but no values or tuning procedure are given.
  • duration loss weight λ = not reported
    L_total = L_CE + λ L_dur in Eq. (10), with λ unspecified.
  • RVQ channel split and codebook sizes = 1 semantic + 5 acoustic channels; sizes not reported
    Architecture choice determining compression and capacity; no codebook sizes, vocab sizes, or bitrates given.
  • inference schedule and sampling settings = [8,4,1,1,1,1] steps; top 10% logits; temperature annealing 0.1→0
    Hand-tuned decode schedule; no sensitivity analysis is provided.
axioms (5)
  • standard math Cross-entropy, KL divergence, and residual vector quantization update rules are valid and implemented correctly.
    These are the mathematical primitives in Equations (1), (9), and the RVQ pipeline; the paper provides no formalization, only standard usage.
  • domain assumption Wav2vec-Bert semantic features are a good distillation target for the 0-th RVQ channel.
    Section III-A; no evidence that this choice doesn't hurt speaker similarity, and the paper concedes codec reconstruction SIM-o is only 0.668.
  • domain assumption Ground-truth phone durations from alignments are accurate and perceptually meaningful.
    Section III-B uses these durations as supervision for L_dur and for placing anchor tokens; no alignment error or human-validation analysis is reported.
  • ad hoc to paper The center anchor at floor(t_i/2) is the best single anchor, and padding tokens elsewhere are harmless.
    Section III-B; the central strategy is motivated intuitively and an ablation supports it, but only on two test sets with no per-case or statistical analysis.
  • ad hoc to paper Masked iterative inference with the specific cosine-bound schedule and 8/4/1/1/1/1 steps gives near-converged quality.
    Section IV inference; the schedule and step counts are chosen empirically with no sensitivity study.

pith-pipeline@v1.3.0-alltime-deepseek · 7830 in / 11067 out tokens · 113199 ms · 2026-08-01T11:28:24.230370+00:00 · methodology

0 comments
read the original abstract

The trade-off between robustness, latency, and prosody critically challenges text-to-speech (TTS) systems. Autoregressive models, despite fidelity, are slow and error-prone; non-autoregressive (NAR) alternatives, while fast, often sacrifice prosodic naturalness via rigid alignments. This paper introduces StellarTTS, a novel mobile-optimized NAR TTS framework based on a sparse temporal embedding strategy, enabling granular control of phoneme duration, pronunciation, and prosody. Furthermore, we propose a semantic-aware codec that facilitates efficient single-stage decoding. Conditioned on the sparse temporal embedding, our 83M-parameter lightweight masked generative transformer achieves a real-time factor (RTF) of 0.08. Experiments demonstrate that StellarTTS attains lower latency and stronger robustness compared to state-of-the-art TTS systems, while maintaining competitive performance in audio quality, prosodic naturalness, and speaker similarity.

Figures

Figures reproduced from arXiv: 2607.19859 by Bing Han, Huiyan Li, Jinling He, Kaicheng Luo, Xiaoyang Xing, Xuefei Gong, Yanmin Qian, Yujie Hou, Yutao Sun.

Figure 1
Figure 1. Figure 1: An overview of StellarTTS. (a) demonstrates the semantic-aware codec, where the 0-th channel of RVQ is distillated with semantic feature extracted [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Impact of Duration Control Strategies on WER: StellarTTS and [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

23 extracted references · 16 linked inside Pith

  1. [1]

    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, Huam- ing Wang, and Jinyu Li, “Neural codec language models are zero-shot text to speech synthesizers,”arXiv preprint arXiv:2301.02111, 2023

    Pith/arXiv arXiv 2023

  2. [2]

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

    Zhihao Du, Qian Chen, Shiliang Zhang, Kai Hu, Heng Lu, Yexin Yang, Hangrui Hu, Siqi Zheng, Yue Gu, and Ziyang Ma, “Cosyvoice: A scalable multilingual zero- shot text-to-speech synthesizer based on supervised se- mantic tokens,”arXiv preprint arXiv:2407.05407, 2024

    Pith/arXiv arXiv 2024

  3. [3]

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

    Zhihao Du, Yuxuan Wang, Qian Chen, Xian Shi, Xiang Lv, Tianyu Zhao, Zhifu Gao, Yexin Yang, Changfeng Gao, and Hui Wang, “ Cosyvoice 2: Scalable streaming speech synthesis with large language models,”arXiv preprint arXiv:2412.10117, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Soundstorm: Efficient parallel audio generation,

    Zal ´an Borsos, Matt Sharifi, Damien Vincent, Eugene Kharitonov, Neil Zeghidour, and Marco Tagliasacchi, “Soundstorm: Efficient parallel audio generation,”arXiv preprint arXiv:2305.09636, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    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, Shunsi Zhang, and Zhizheng Wu, “Maskgct: Zero-shot text-to- speech with masked generative codec transformer,”arXiv preprint arXiv:2409.00750, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    E2 tts: Embarrass- ingly easy fully non-autoregressive zero-shot tts,

    Sefik Emre Eskimez, Xiaofei Wang, Manthan Thakker, Canrun Li, Chung-Hsien Tsai, Zhen Xiao, Hemin Yang, Zirun Zhu, Min Tang, and Xu Tan, “E2 tts: Embarrass- ingly easy fully non-autoregressive zero-shot tts,”arXiv preprint arXiv:2406.18009, 2024

    Pith/arXiv arXiv 2024

  7. [7]

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

    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,”arXiv preprint arXiv:2410.06885, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Naturalspeech 3: Zero- shot speech synthesis with factorized codec and diffusion models,

    Zeqian Ju, Yuancheng Wang, Kai Shen, Xu Tan, De- tai Xin, Dongchao Yang, Yanqing Liu, Yichong Leng, Kaitao Song, and Siliang Tang, “Naturalspeech 3: Zero- shot speech synthesis with factorized codec and diffusion models,”arXiv preprint arXiv:2403.03100, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Fastspeech: Fast, robust and controllable text to speech,

    Yi Ren, Yangjun Ruan, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu, “Fastspeech: Fast, robust and controllable text to speech,” inProceedings of the International Conference on Neural Information Process- ing Systems, 2019, pp. 3165–3174

    2019

  10. [10]

    Fastspeech 2: Fast and high-quality end-to-end text to speech,

    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 preprint arXiv:2006.04558, 2020

    Pith/arXiv arXiv 2006

  11. [11]

    V oicebox: Text-guided multilingual universal speech generation at scale,

    Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, and Jay Mahadeokar, “V oicebox: Text-guided multilingual universal speech generation at scale,” inProceedings of the International Conference on Neural Information Processing Systems, 2023, pp. 14005-14034

    2023

  12. [12]

    Naturalspeech 2: Latent diffusion models are natural and zero-shot speech and singing synthesizers,

    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,”arXiv preprint arXiv:2304.09116, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Llama: Open and efficient founda- tion language models,

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth ´ee Lacroix, Bap- tiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample, “Llama: Open and efficient founda- tion language models,”arXiv preprint arXiv:2302.13971, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    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, and Lu Gao, “Seed-tts: A family of high-quality versatile speech generation models,”arXiv preprint arXiv:2406.02430, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Neural discrete representation learning,

    Aaron Van Den Oord and Oriol Vinyals, “Neural discrete representation learning,” inProceedings of the Inter- national Conference on Neural Information Processing Systems, 2017, pp. 6309–6318

    2017

  16. [16]

    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, vol. 30, pp. 495-507, 2021

    2021

  17. [17]

    High fidelity neural audio compression,

    Alexandre D ´efossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi, “High fidelity neural audio compression,” arXiv preprint arXiv:2210.13438, 2022

    Pith/arXiv arXiv 2022

  18. [18]

    Hubert: Self-supervised speech repre- sentation learning by masked prediction of hidden units,

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrah- man Mohamed, “Hubert: Self-supervised speech repre- sentation learning by masked prediction of hidden units,” IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2022

    2022

  19. [19]

    Wavlm: Large-scale self- supervised pretraining for full stack speech processing,

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, and Xiong Xiao, “Wavlm: Large-scale self- supervised pretraining for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505-1518, 2022

    2022

  20. [20]

    W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,

    Yu-An Chung, Yu Zhang, Wei Han, Chung-Cheng Chiu, James Qin, Ruoming Pang, and Yonghui Wu, “W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,” in 2021 IEEE Automatic Speech Recognition and Under- standing Workshop (ASRU). IEEE, 2021, pp. 244–250

    2021

  21. [21]

    3D-Speaker: A Large-Scale Multi- Device, Multi-Distance, and Multi-Dialect Corpus for Speech Representation Disentanglement,

    Siqi Zheng, Luyao Cheng, Yafeng Chen, Hui Wang, and Qian Chen, “3D-Speaker: A Large-Scale Multi- Device, Multi-Distance, and Multi-Dialect Corpus for Speech Representation Disentanglement,”arXiv preprint arXiv:2306.15354, 2023

    Pith/arXiv arXiv 2023

  22. [22]

    Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech gener- ation,

    Haorui He, Zengqiang Shang, Chaoren Wang, Xuyuan Li, Yicheng Gu, Hua Hua, Liwei Liu, Chen Yang, Jiaqi Li, and Peiyang Shi, “Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech gener- ation,”arXiv preprint arXiv:2407.05361, 2024

    Pith/arXiv arXiv 2024

  23. [23]

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

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

    Pith/arXiv arXiv 2024