pith. sign in

arxiv: 2606.09141 · v2 · pith:M4EMYDIEnew · submitted 2026-06-08 · 📡 eess.AS · cs.SD

FlashTTS: Fast Streaming TTS with MTP Acceleration and X-pred Mean Flow Distillation

Hanke Xie , Xiaming Ren , Dake Guo , Ruonan You , Wenhao Li , Jingbin Hu , Guobin Ma , Huakang Chen
show 5 more authors
Kejie Xu Rui Huang Weiguo Tan Xianrong Wang Lei Xie
This is my paper

Pith reviewed 2026-06-27 15:13 UTC · model grok-4.3

classification 📡 eess.AS cs.SD
keywords streaming TTSlow latencymulti-token predictionflow matchingfirst-packet latencyvoice cloningspeech dialogue systems
0
0 comments X

The pith

FlashTTS reaches 325 ms first-packet latency in streaming TTS through a lagged multi-track design paired with multi-token prediction and two-step flow matching.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents FlashTTS to satisfy the low-latency and streaming requirements of modern speech dialogue systems. Existing single-codebook LLM-based TTS methods use multi-stage pipelines that force sentence-level buffering and incur high end-to-end latency from autoregressive steps and multi-step flow matching. FlashTTS replaces these with a native streaming architecture that handles text and speech inputs on the fly. It accelerates acoustic generation by running parallel multi-token prediction together with an X-pred mean flow matching decoder that finishes in exactly two function evaluations. Experiments report the resulting first-packet latency of 325 ms alongside retained zero-shot voice cloning and cross-lingual intelligibility.

Core claim

FlashTTS introduces a lagged multi-track architecture that natively processes streaming text and speech inputs without sentence-level buffering, integrates parallel Multi-Token Prediction with an X-pred mean flow matching decoder to achieve high-fidelity token-to-mel generation in two function evaluations, and thereby reduces first-packet latency to 325 ms while preserving zero-shot voice cloning and cross-lingual performance.

What carries the argument

The lagged multi-track architecture that processes streaming inputs, accelerated by parallel Multi-Token Prediction (MTP) and an X-pred mean flow matching decoder that completes acoustic generation in two function evaluations.

If this is right

  • Real-time speech dialogue systems can operate with sub-second response times without sentence buffering.
  • Streaming TTS becomes compatible with zero-shot voice cloning and cross-lingual generation in a single model.
  • End-to-end latency drops by removing multi-stage pipelines and multi-step flow matching.
  • Open-source release of model code and checkpoints allows direct integration into dialogue applications.

Where Pith is reading between the lines

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

  • The two-evaluation flow matching step may extend to other token-to-continuous generation tasks that currently require many denoising steps.
  • Joint optimization of input buffering and decoder speed could reduce memory footprint on resource-constrained devices.
  • The same lagged multi-track pattern might apply to streaming generation in other modalities such as music or video.
  • If the 2-NFE decoder generalizes, it offers a route to lower compute budgets for on-device TTS.

Load-bearing premise

The lagged multi-track architecture together with multi-token prediction and the X-pred mean flow matching decoder can be jointly trained to generate high-fidelity mel-spectrograms without quality loss detectable in listening tests or downstream ASR metrics.

What would settle it

A side-by-side listening test or ASR word-error-rate comparison on the same zero-shot voices that shows clear degradation in naturalness or intelligibility relative to non-streaming baselines.

Figures

Figures reproduced from arXiv: 2606.09141 by Dake Guo, Guobin Ma, Hanke Xie, Huakang Chen, Jingbin Hu, Kejie Xu, Lei Xie, Rui Huang, Ruonan You, Weiguo Tan, Wenhao Li, Xiaming Ren, Xianrong Wang.

Figure 1
Figure 1. Figure 1: Architecture Overview of FlashTTS: (a) Stage 1 Train￾ing of the Stacked Inputs Track Structure. (b) Stage 2 MTP Training guage models. VALL-E [3] represents a pioneering system that leverages an LLM operating on discrete tokens with a residual vector quantization (RVQ) tokenizer [4]. In this de￾sign, tokens from the first codebook are predicted autoregres￾sively, while tokens from the remaining codebooks a… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of different input schemes. that combine LLM-based with flow matching speech synthe￾sis. Deploying such interactive systems requires both extremely low latency and the capability to stream inputs and outputs. However, conventional TTS systems do not natively support streaming input. Although recent work [8, 21] introduces in￾terleaved generation to reduce sentence-level waiting time, the overall… view at source ↗
read the original abstract

Recent progress in speech dialogue systems requires Text-to-Speech (TTS) models to be faster and more responsive. Modern speech dialogue systems impose two primary requirements on TTS models: low latency and support for streaming inputs and outputs. However, most existing single-codebook LLM-based TTS methods rely on multi-stage pipelines that lack native streaming capabilities. These systems typically suffer from high end-to-end latency due to slow autoregressive prediction and multi-step flow matching. To address these limitations, we propose FlashTTS, an open-source and low-latency streaming TTS framework. FlashTTS introduces a lagged multi-track architecture that natively processes streaming text and speech inputs, thereby eliminating the need for sentence-level buffering. To accelerate acoustic generation, we integrate parallel Multi-Token Prediction (MTP) with an X-pred mean flow matching decoder. This configuration achieves high-fidelity token-to-mel generation in exactly two function evaluations (2-NFE). By jointly optimizing input processing and decoding efficiency, FlashTTS offers a practical foundation for real-time speech dialogue systems. Experiments show that FlashTTS substantially reduces First-Packet Latency to 325ms compared to robust streaming baselines, all while preserving strong zero-shot voice cloning and cross-lingual intelligibility. Speech samples are available. The model code and checkpoints will be released as open source.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces FlashTTS, an open-source streaming TTS framework using a lagged multi-track architecture for native processing of streaming text and speech inputs, combined with parallel Multi-Token Prediction (MTP) and an X-pred mean flow matching decoder restricted to exactly 2 function evaluations (2-NFE). It claims this yields high-fidelity mel-spectrogram generation, reducing First-Packet Latency to 325 ms versus robust streaming baselines while preserving zero-shot voice cloning and cross-lingual intelligibility, with code and checkpoints to be released.

Significance. If the latency and quality claims are substantiated with rigorous experiments, the work could offer a practical contribution to real-time speech dialogue systems by enabling lower end-to-end latency without multi-stage pipelines. The explicit commitment to open-sourcing the model code and checkpoints is a strength that supports reproducibility.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (Experiments): The central claim of a 325 ms First-Packet Latency reduction is presented without any quantitative tables, error bars, dataset specifications, baseline configurations, or ablation results on the 2-NFE regime, making it impossible to verify whether the comparison to streaming baselines is robust or affected by post-hoc selection.
  2. [§3] §3 (Method): The assertion that the lagged multi-track architecture, MTP, and X-pred mean flow matching decoder can be jointly trained to produce high-fidelity mels without quality degradation (required for the downstream zero-shot cloning and intelligibility claims) lacks any training procedure details, loss curves, MOS scores, WER metrics, or ablation on the distillation process.
minor comments (2)
  1. The abstract states that 'Speech samples are available' and code will be released, but no URLs or repository links are provided in the text.
  2. [§3] Notation for 'X-pred mean flow distillation' is introduced without a clear definition or equation relating it to standard flow matching.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive feedback. We appreciate the emphasis on rigorous experimental validation and will revise the manuscript to strengthen the presentation of results and methods.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Experiments): The central claim of a 325 ms First-Packet Latency reduction is presented without any quantitative tables, error bars, dataset specifications, baseline configurations, or ablation results on the 2-NFE regime, making it impossible to verify whether the comparison to streaming baselines is robust or affected by post-hoc selection.

    Authors: We agree that the current version of the abstract and §4 would benefit from expanded quantitative support. In the revised manuscript we will add tables reporting First-Packet Latency with error bars obtained from multiple runs, explicit dataset specifications, baseline configurations (including their streaming parameters), and dedicated ablations isolating the 2-NFE regime. These additions will make the 325 ms comparison fully transparent and reproducible. revision: yes

  2. Referee: [§3] §3 (Method): The assertion that the lagged multi-track architecture, MTP, and X-pred mean flow matching decoder can be jointly trained to produce high-fidelity mels without quality degradation (required for the downstream zero-shot cloning and intelligibility claims) lacks any training procedure details, loss curves, MOS scores, WER metrics, or ablation on the distillation process.

    Authors: We acknowledge that §3 currently omits sufficient training and evaluation details. The revised version will expand this section with the full joint-training procedure, loss curves, MOS and WER results, and ablations on the X-pred mean-flow distillation process to demonstrate that high-fidelity mel generation is preserved for the downstream zero-shot and cross-lingual claims. revision: yes

Circularity Check

0 steps flagged

No circularity; performance claims are measured outcomes, not reductions to fitted inputs

full rationale

The paper presents an architectural proposal (lagged multi-track + parallel MTP + X-pred mean-flow decoder at 2 NFE) whose headline metrics (325 ms FPL, preserved zero-shot cloning and intelligibility) are reported as experimental results rather than quantities derived by construction from the same fitted parameters or self-citations. No equations, loss definitions, or parameter-fitting steps are shown that would make the latency or quality numbers tautological with the model definition itself. The central claims therefore remain externally falsifiable via listening tests or ASR WER and do not reduce to the inputs by the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central performance claims rest on standard supervised training of a neural TTS model plus the unstated assumption that the new architectural components can be optimized jointly without introducing instability or quality loss; no new physical constants or external benchmarks are invoked.

free parameters (2)
  • number of lagged tracks
    Chosen to balance streaming responsiveness against context length; value not stated in abstract.
  • MTP prediction horizon
    Hyperparameter controlling how many future tokens are predicted in parallel.
axioms (1)
  • domain assumption Flow matching can be distilled to exactly two function evaluations while retaining mel-spectrogram fidelity sufficient for downstream vocoding.
    Invoked to justify the 2-NFE claim; no derivation supplied in abstract.
invented entities (1)
  • X-pred mean flow distillation no independent evidence
    purpose: Accelerate acoustic token-to-mel generation to two network evaluations.
    Presented as a novel component; no independent evidence outside the paper is referenced.

pith-pipeline@v0.9.1-grok · 5813 in / 1442 out tokens · 18178 ms · 2026-06-27T15:13:07.770193+00:00 · methodology

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

40 extracted references · 29 canonical work pages · 13 internal anchors

  1. [1]

    Introduction Building on the generative capacity of large language models (LLMs), modern text-to-speech (TTS) systems have reached a stage where synthesized speech is nearly indistinguishable from human speech. By imitating the timbre, prosody, and speaking style of reference audio, these systems achieve high naturalness and strong zero-shot voice cloning...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    Model Architecture In this section, we present FlashTTS, a low-latency TTS model explicitly designed for streaming input scenarios

    FlashTTS 2.1. Model Architecture In this section, we present FlashTTS, a low-latency TTS model explicitly designed for streaming input scenarios. As illus- trated in Figure 1, the model architecture and training pipeline are structured into two primary stages. Stage 1 (Figure 1a) establishes the core generation pathway, wherein a trainable decoder-only tr...

  3. [3]

    Datasets FlashTTS is trained on approximately 300,000 hours of open- source speech data, comprising Emilia, Emilia-Yodas [29, 30, 31], LibriHeavy, and WenetSpeech4TTS

    Experimental Setup 3.1. Datasets FlashTTS is trained on approximately 300,000 hours of open- source speech data, comprising Emilia, Emilia-Yodas [29, 30, 31], LibriHeavy, and WenetSpeech4TTS. We evaluate zero-shot performance on the Seed-TTS3 and MiniMax4 multilingual test sets. 3https://github.com/BytedanceSpeech/seed-tts-eval 4https://huggingface.co/dat...

  4. [4]

    Experimental Results 4.1. Latency and Quality Analysis All latency evaluations are conducted on the Minimax sub- set using a single NVIDIA RTX 4090 GPU, with the tex- 5https://github.com/modelscope/FunASR Table 2:Objective evaluation metrics on the multilingual test set. Lower WER and higher SIM indicate better performance. “–” denotes unsupported languag...

  5. [5]

    Its multi-track formulation natively processes streaming inputs, completely circumventing traditional sentence-level buffering

    Conclusion We propose FlashTTS, a low-latency LLM-based streaming TTS framework addressing the high latency bottleneck in real- time dialogue systems via a lagged multi-track architecture, par- allel MTP, and an X-pred Mean Flow decoder. Its multi-track formulation natively processes streaming inputs, completely circumventing traditional sentence-level bu...

  6. [6]

    They are not used to generate any core content, research ideas, experimental designs, results, or major textual parts of the paper

    Generative AI Use Disclosure Generative AI tools are used in this work only for language edit- ing, polishing, and formatting of the manuscript. They are not used to generate any core content, research ideas, experimental designs, results, or major textual parts of the paper. All scien- tific contributions, including model design, experiments, analy- sis,...

  7. [7]

    Neural discrete represen- tation learning,

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

    2017

  8. [8]

    Finite Scalar Quantization: VQ-VAE Made Simple

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

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

    Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers

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

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  10. [10]

    High fidelity neural audio compression,

    A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,”Trans. Mach. Learn. Res., vol. 2023, 2023

    2023

  11. [11]

    V oicecraft: Zero-shot speech editing and text-to-speech in the wild,

    P. Peng, P.-Y . Huang, S.-W. Li, A. Mohamed, and D. Harwath, “V oicecraft: Zero-shot speech editing and text-to-speech in the wild,” inProceedings of the 62nd Annual Meeting of the Asso- ciation for Computational Linguistics (V olume 1: Long Papers), 2024, pp. 12 442–12 462

    2024

  12. [12]

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

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

    work page arXiv 2024

  13. [13]

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

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

    work page arXiv 2024

  14. [14]

    Fireredtts-2: Towards long conversational speech generation for podcast and chatbot,

    K. Xie, F. Shen, J. Li, F. Xie, X. Tang, and Y . Hu, “Fireredtts-2: Towards long conversational speech generation for podcast and chatbot,”arXiv preprint arXiv:2509.02020, 2025

    work page arXiv 2025

  15. [15]

    Qwen3-TTS Technical Report

    H. Hu, X. Zhu, T. He, D. Guo, B. Zhang, X. Wang, Z. Guo, Z. Jiang, H. Hao, Z. Guoet al., “Qwen3-tts technical report,” arXiv preprint arXiv:2601.15621, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    Quarkaudio technical report,

    C. Liu, H. Yan, S. Xue, X. Liang, X. Chen, B. Gong, Z. Xue, and G. Song, “Quarkaudio technical report,”arXiv preprint arXiv:2512.20151, 2025

    work page arXiv 2025

  17. [17]

    Better speech synthesis through scaling,

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

    work page arXiv 2023

  18. [18]

    Sac: Neural speech codec with semantic-acoustic dual-stream quantization,

    W. Chen, X. Wang, R. Yan, Y . Chen, Z. Niu, Z. Ma, X. Li, Y . Liang, H. Wen, S. Yinet al., “Sac: Neural speech codec with semantic-acoustic dual-stream quantization,”arXiv preprint arXiv:2510.16841, 2025

    work page arXiv 2025

  19. [19]

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

    Z. Ye, X. Zhu, C.-M. Chan, X. Wang, X. Tan, J. Lei, Y . Peng, H. Liu, Y . Jin, Z. Daiet al., “Llasa: Scaling train-time and inference-time compute for llama-based speech synthesis,”arXiv preprint arXiv:2502.04128, 2025

    work page arXiv 2025

  20. [20]

    Spark-TTS: An Efficient LLM-Based Text-to-Speech Model with Single-Stream Decoupled Speech Tokens

    X. Wang, M. Jiang, Z. Ma, Z. Zhang, S. Liu, L. Li, Z. Liang, Q. Zheng, R. Wang, X. Fenget al., “Spark-tts: An efficient llm- based text-to-speech model with single-stream decoupled speech tokens,”arXiv preprint arXiv:2503.01710, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

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

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

    work page arXiv 2024

  22. [22]

    Seed-TTS: A Family of High-Quality Versatile Speech Generation Models

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  23. [23]

    CosyVoice: A Scalable Multilingual Zero-shot Text-to-speech Synthesizer based on Supervised Semantic Tokens

    Z. Du, Q. Chen, S. Zhang, K. Hu, H. Lu, Y . Yang, H. Hu, S. Zheng, Y . Gu, Z. Ma, Z. Gao, and Z. Yan, “Cosyvoice: A scal- able multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens,”CoRR, vol. abs/2407.05407, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  24. [24]

    CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Models

    Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wang, F. Yu, H. Liu, Z. Sheng, Y . Gu, C. Deng, W. Wang, S. Zhang, Z. Yan, and J. Zhou, “Cosyvoice 2: Scalable streaming speech synthesis with large language models,”CoRR, vol. abs/2412.10117, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  25. [25]

    Glm-tts technical report,

    J. Cui, Z. Yang, N. Li, J. Tian, X. Ma, Y . Zhang, G. Chen, R. Yang, Y . Cheng, Y . Zhou, G. Yu, X. Gu, and J. Tang, “Glm-tts technical report,” 2025. [Online]. Available: https://arxiv.org/abs/2512.14291

    work page arXiv 2025

  26. [26]

    X-talk: On the underestimated potential of modular speech-to-speech dialogue system,

    Z. Liu, Y . Duan, M. Wang, P. Feng, H. Zhang, X. Xing, Y . Shan, H. Zhu, Y . Dai, C. Luet al., “X-talk: On the underestimated potential of modular speech-to-speech dialogue system,”arXiv preprint arXiv:2512.18706, 2025

    work page arXiv 2025

  27. [27]

    Soulx-podcast: Towards realis- tic long-form podcasts with dialectal and paralinguistic diversity,

    H. Xie, H. Lin, W. Cao, D. Guo, W. Tian, J. Wu, H. Wen, R. Shang, H. Liu, Z. Jianget al., “Soulx-podcast: Towards realis- tic long-form podcasts with dialectal and paralinguistic diversity,” arXiv preprint arXiv:2510.23541, 2025

    work page arXiv 2025

  28. [28]

    Mean Flows for One-step Generative Modeling

    Z. Geng, M. Deng, X. Bai, J. Z. Kolter, and K. He, “Mean flows for one-step generative modeling,”arXiv preprint arXiv:2505.13447, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  29. [29]

    Back to Basics: Let Denoising Generative Models Denoise

    T. Li and K. He, “Back to basics: Let denoising generative models denoise,”arXiv preprint arXiv:2511.13720, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    DeepSeek-V3 Technical Report

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruanet al., “Deepseek-v3 technical re- port,”arXiv preprint arXiv:2412.19437, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  31. [31]

    Llasa+: Free lunch for accelerated and streaming llama-based speech syn- thesis,

    W. Tian, X. Zhu, H. Xie, Z. Ye, W. Xue, and L. Xie, “Llasa+: Free lunch for accelerated and streaming llama-based speech syn- thesis,”arXiv preprint arXiv:2508.06262, 2025

    work page arXiv 2025

  32. [32]

    Meanvc: Lightweight and streaming zero-shot voice conversion via mean flows,

    G. Ma, J. Yao, Z. Ning, Y . Jiang, L. Xiong, L. Xie, and P. Zhu, “Meanvc: Lightweight and streaming zero-shot voice conversion via mean flows,”arXiv preprint arXiv:2510.08392, 2025

    work page arXiv 2025

  33. [33]

    Int- meanflow: Few-step speech generation with integral velocity dis- tillation,

    W. Wang, R. Cao, Y . Guo, Z. Chen, K. Chen, and Y . Huo, “Int- meanflow: Few-step speech generation with integral velocity dis- tillation,”arXiv preprint arXiv:2510.07979, 2025

    work page arXiv 2025

  34. [34]

    Streamflow: Streaming flow matching with block-wise guided attention mask for speech token decoding,

    D. Guo, J. Yao, L. Ma, H. Wang, and L. Xie, “Streamflow: Streaming flow matching with block-wise guided attention mask for speech token decoding,” inNational Conference on Man- Machine Speech Communication. Springer, 2025, pp. 75–86

    2025

  35. [35]

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

    H. He, Z. Shang, C. Wang, X. Li, Y . Gu, H. Hua, L. Liu, C. Yang, J. Li, P. Shiet al., “Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 885–890

    2024

  36. [36]

    Wenetspeech4tts: A 12,800-hour mandarin tts corpus for large speech generation model bench- mark,

    L. Ma, D. Guo, K. Song, Y . Jiang, S. Wang, L. Xue, W. Xu, H. Zhao, B. Zhang, and L. Xie, “Wenetspeech4tts: A 12,800-hour mandarin tts corpus for large speech generation model bench- mark,”arXiv preprint arXiv:2406.05763, 2024

    work page arXiv 2024

  37. [37]

    Libriheavy: A 50,000 hours asr corpus with punc- tuation casing and context,

    W. Kang, X. Yang, Z. Yao, F. Kuang, Y . Yang, L. Guo, L. Lin, and D. Povey, “Libriheavy: A 50,000 hours asr corpus with punc- tuation casing and context,” inICASSP 2024-2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 991–10 995

    2024

  38. [38]

    Qwen Technical Report

    J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. Ge, Y . Han, F. Huanget al., “Qwen technical report,”arXiv preprint arXiv:2309.16609, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  39. [39]

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

    Y . Wang, H. Zhan, L. Liu, R. Zeng, H. Guo, J. Zheng, Q. Zhang, X. Zhang, S. Zhang, and Z. Wu, “Maskgct: Zero-shot text-to- speech with masked generative codec transformer,”arXiv preprint arXiv:2409.00750, 2024

    work page arXiv 2024

  40. [40]

    F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024