Pith. sign in

REVIEW 4 major objections 5 minor 38 references

Accelerating Flow-Matching-Based Text-to-Speech via Empirically Pruned Step Sampling

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read By keeping early steps dense and pruning later steps, a seven-step schedule runs F5-TTS four times faster with comparable speech quality.

desk verdict Genuinely useful empirical trick: pruning later flow-matching steps gives a real 4x speedup on F5-TTS with quality intact, and the E2 transfer is credible, but the headline schedule was tuned on the same test set it is evaluated on. read the letter →

arxiv 2505.19931 v2 pith:73FN3DFN submitted 2025-05-26 eess.AS cs.SD

Qixi Zheng , Yushen Chen , Zhikang Niu , Ziyang Ma , Xiaofei Wang , Kai Yu , Xie Chen This is my paper
classification eess.AScs.SD
keywords flowmatchingtext-to-speechinferenceefficiencystepschedulingnon-uniformsamplingtraining-freeaccelerationF5-TTSE2TTS
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that most of the sampling steps in flow-matching text-to-speech models are redundant, and that a hand-designed non-uniform time schedule can remove them without retraining. The authors inspect the sampling trajectory of F5-TTS and observe that the later phase is nearly linear, so large jumps are safe there, while the early phase is curved and needs fine steps. They propose Empirically Pruned Step Sampling (EPSS), a schedule that keeps steps at small times and prunes them at large times. With EPSS, F5-TTS generates speech in 7 function evaluations instead of 32, cutting the real-time factor from 0.123 to 0.030 on an RTX 3090 while keeping word error rate, speaker similarity, and naturalness close to the baseline. The same schedule also works for E2 TTS, suggesting the recipe may apply to other flow-matching models.

What carries the argument

The central object is EPSS (Empirically Pruned Step Sampling), a non-uniform time-step schedule that keeps the early portion of the integration interval dense and prunes the later portion. It is derived from a qualitative PCA-based inspection of the F5-TTS sampling trajectory, which shows the trajectory is curved early and nearly linear late. EPSS does not modify the network or the ODE solver; it only changes the set of time points at which the vector field is evaluated, so it is training-free and plug-and-play.

What would settle it

Compute a curvature or deviation-from-straight-line statistic for the second half of the sampling trajectory over a diverse set of utterances, speakers, and languages; if a substantial fraction of trajectories show strong curvature late, the dense-early/sparse-late schedule will underperform on those inputs and the reported 7-NFE quality will not generalize.

Watch

Extended reading notes

Core claim

The central discovery is an empirical property of FM-OT-based TTS sampling: the inference trajectory is not uniform in difficulty. Early steps operate on near-noise inputs and exhibit strong curvature, so they must be sampled finely; later steps move along a nearly straight line and tolerate very large step sizes. EPSS encodes this by using a 7-step index pattern $\{0, 2, 4, 6, 8, 16, 24, 32\}$ on the 32-step grid (then passing it through the sway sampling function), which matches the quality of the 32-step baseline while cutting computation by 78%. The paper also shows that pruning a later step (at $t=3/4$) hurts far less than pruning an early step (at $t=3/16$), and that the schedule transfers to E2 TTS.

Load-bearing premise

The load-bearing premise is that the later phase of the F5-TTS sampling trajectory is nearly linear for typical utterances and so tolerates large step sizes, a claim supported only by a qualitative PCA of one example rather than a quantitative measure or theory.

Editorial extensions

If this is right

  • Fast F5-TTS reaches a real-time factor of 0.030 on an RTX 3090, about 4x faster than the 32-step F5-TTS baseline (RTF 0.123), while keeping WER, speaker similarity, and UTMOS close to the baseline.
  • Because EPSS only changes the time-step schedule, it is training-free and can be combined with sway sampling and, in principle, with distillation or other accelerators without retraining the model.
  • At 16 NFE the pruned schedule improves on the 32-step baseline in all reported metrics, so the benefit is not limited to the low-NFE regime.
  • The same 7-NFE EPSS schedule applied to E2 TTS restores most of the quality lost by naive 7-step sampling, showing the schedule is not specific to F5-TTS.
  • There is a practical floor around 6 NFE: 5-step schedules degrade sharply (WER rises and UTMOS drops), indicating how far empirical pruning can go before quality collapses.

Reading between the lines

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

  • A quantitative linearity diagnostic (e.g., measuring the angle or residual between consecutive velocity vectors over many utterances) would turn the single-example PCA observation into a testable property and could predict which inputs the 7-step schedule will struggle with.
  • If the late-phase linearity holds broadly, EPSS-style schedules could be tuned per model on a small validation set, making the hand-picked index pattern a special case of a more general step-allocation rule.
  • The same dense-early/sparse-late idea might apply to other flow-matching generative models (images, video, audio beyond speech), but only if their trajectories share the curvature profile; measuring that is a direct next experiment.
  • Because EPSS is schedule-only, it could be layered with distillation methods: distillation straightens the trajectory further, which would make even the early steps more prunable and potentially push the step count below five.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Empirically Pruned Step Sampling (EPSS), a training-free non-uniform time-step schedule for flow-matching-based TTS, and evaluates it on F5-TTS and E2 TTS. EPSS keeps early sampling steps dense and prunes later steps, achieving 7-NFE synthesis with an RTF of 0.030 on an RTX 3090, about 4x faster than the original 32-NFE F5-TTS, while reporting comparable WER, SIM-o, and UTMOS on LibriSpeech-PC test-clean and Seed-TTS-eval. The schedule is presented in Table 2 with ablations in Table 3, and the method is also applied to E2 TTS, where it improves over the 7-NFE uniform baseline.

Significance. If the reported results are robust, EPSS is a simple, plug-and-play acceleration method with a clear practical benefit for existing flow-matching TTS systems, requiring no retraining. The paper provides a concrete schedule with exact time-step indices, a reproducible RTF measurement protocol, and a direct transfer test to a second backbone model (E2 TTS), which is a valuable strength. The analysis of the sampling trajectory, while qualitative, motivates a pruning pattern that is empirically validated by the 6a/6b ablation. However, the significance is tempered by the lack of statistical error bars and the fact that the schedule was tuned on the same test set used for the headline LibriSpeech-PC results, leaving the magnitude of the true quality-speed trade-off uncertain.

major comments (4)
  1. [Section 4.4 / Tables 1 and 3] The EPSS time-step schedule was selected using ablations on the LibriSpeech-PC test-clean set (Table 3), and the same set is then used for the headline results in Table 1. This means the reported WER of 2.45, SIM-o of 0.66, and UTMOS of 3.84 for the 7-step EPSS configuration may reflect selection over the explored grid rather than the performance of a fixed, pre-specified schedule. Please perform the schedule search on a held-out validation split (e.g., a portion of the training or development data) and evaluate the chosen schedule on LibriSpeech-PC test-clean only once, or alternatively report the Seed-TTS-eval results as the primary unbiased evidence and explicitly state the size of the grid searched (e.g., how many 7-step schedules were tried before settling on [0,2,4,6,8,16,24,32]).
  2. [Table 1] The paper reports all metrics as averages over three random seeds but provides no standard deviations, confidence intervals, or significance tests. For the central claim of 'comparable performance,' the relevant differences are small: F5-TTS 32-step WER is 2.37 versus 2.45 for EPSS 7-step on LibriSpeech-PC, and 1.70 versus 1.74 on Seed-TTS test-en. Without error bars, it is impossible to determine whether these differences are meaningful or noise. Please include per-seed results or error bars, and perform a paired significance test (e.g., a bootstrap or Wilcoxon test) for the key comparisons between EPSS at 7 NFE and the baselines at 7, 16, and 32 NFE.
  3. [Section 3.1 and Figure 1-2] The theoretical justification for EPSS rests on the claim that the later phase of the F5-TTS sampling trajectory is nearly linear and therefore tolerates large step sizes, while the early phase is nonlinear and requires fine steps. This is supported only by a qualitative PCA visualization of a single frame (Figures 1 and 2), which the paper itself calls 'qualitative analysis.' Since this premise is used to derive the dense-early/sparse-late pattern, please provide a quantitative measure of trajectory linearity or curvature (e.g., the deviation of the trajectory from a straight line in the latent space, or the magnitude of the second derivative of the flow, averaged over many utterances) and show that the chosen step sizes respect the estimated linear region. The 6a/6b ablation in Table 3 is a good start, but it only compares two pruning points and does not establish that the specific indices in Table 2 are robust or uniquely determined.
  4. [Section 4.3 / E2 TTS transfer] The E2 TTS results apply the EPSS schedule selected for F5-TTS without any adaptation, which is a reasonable generalization test, but the paper should explicitly acknowledge that the schedule was not tuned for E2 TTS and that the reported E2 TTS numbers are therefore a lower bound on what a model-specific schedule might achieve. This clarification is important because the abstract's claim of 'strong generalization ability' currently rests on a single fixed schedule applied to one additional model, and the reader might otherwise infer that the schedule was adapted.
minor comments (5)
  1. [Section 4.2] The evaluation section lists two ASR systems (Whisper-large-V3 and Paraformer-zh) but does not specify which is used for the English versus Chinese test sets; please clarify the assignment for LibriSpeech-PC test-clean and Seed-TTS test-zh.
  2. [Figures 1 and 2] The PCA plots lack axis labels and an explanation of what the two principal components represent; please add labels (e.g., 'PC1' and 'PC2') and state how many frames from how many utterances are shown.
  3. [Abstract and Conclusion] The phrase 'strong generalization ability' overstates what is demonstrated by applying EPSS to one additional model (E2 TTS); consider softening to 'generalizes to other flow-matching TTS models' or similar.
  4. [Section 4.1 and footnote] The statement 'We will release our updated code ... on the existing F5-TTS GitHub repository' is vague; please provide a specific link to a new repository or a branch where the EPSS implementation will be available.
  5. [Section 4.4] The sentence 'performance sharply declines below 6 NFE' is supported by the 5-NFE UTMOS of 2.91, but WER at 5 NFE (2.55) is actually better than at 7 NFE (2.45), so the decline is not uniform across metrics; consider describing the metric-specific behavior.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EPSS is an empirically selected time-step schedule independently evaluated and transferred to E2 TTS.

full rationale

EPSS is an empirically selected time-step schedule, not a predicted quantity derived from equations that already contain the result. The 7-NFE configuration is defined explicitly in Table 2 and its quality is measured by external metrics in Table 1; no fitted parameter is later renamed as a prediction. The F5-TTS baseline and sway sampling are cited from the authors' own prior work, but the sway sampling formula is stated in the paper and the comparison is a controlled ablation against the 32-NFE baseline, so the self-citation is not load-bearing. The E2 TTS transfer is an independent model with the same schedule and provides additional support. The schedule was explored on LibriSpeech-PC test-clean before being reported on that set, which is a test-set-reuse concern for generalization, but it does not make the central claim true by construction or reduce any equation to its input. Overall, no significant circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The central method adds one empirical knob, the EPSS schedule, which is hand-selected. Background flow-matching math from [13] and F5-TTS sway sampling from [11] are standard inputs. The key domain assumption is that the later trajectory is linear enough to prune steps, which is only qualitatively demonstrated for one example.

free parameters (1)
  • EPSS 7-step schedule indices = After sway re-mapping: {0, 1/16, 1/8, 3/16, 1/4, 1/2, 3/4, 1}; before sway: {0,2,4,6,8,16,24,32}/32
    The exact grid is hand-selected after trajectory inspection and ablation experiments on LibriSpeech-PC test-clean, so it is the main empirical knob of the method.
assumptions (4)
  • standard math Flow matching ODE (Equation 1) accurately models F5-TTS inference and Euler discretization error is governed by vector field curvature.
    Used in Sections 2.1 and 2.2 as background; the paper relies on the standard flow matching formulation.
  • domain assumption Sway sampling is a valid and beneficial time re-mapping for F5-TTS.
    Adopted from the authors' prior F5-TTS work [11] as the baseline schedule; EPSS is defined on top of the sway grid.
  • ad hoc to paper The later-phase trajectory is approximately linear for typical utterances, so large step sizes are safe.
    Section 3.1 justifies pruning late steps using qualitative PCA trajectory plots, but no quantitative linearity measure or error analysis is provided.
  • ad hoc to paper A fixed schedule selected on one model and test set transfers to other utterances, speakers, languages, and flow-matching models.
    The paper applies the F5-TTS-derived schedule to E2 TTS and to all test sets; only E2 TTS is tested as a transfer target, so broader transfer remains assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Accelerating Flow-Matching-Based Text-to-Speech via Empirically Pruned Step Sampling." pith.science (2026). https://pith.science/paper/73FN3DFN

@misc{pith2026250519931,
  author       = {Pith},
  title        = {Pith review of: Accelerating Flow-Matching-Based Text-to-Speech via Empirically Pruned Step Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/73FN3DFN}},
  note         = {Machine review of arXiv:2505.19931}
}
read the original abstract

Flow-matching-based text-to-speech (TTS) models, such as Voicebox, E2 TTS, and F5-TTS, have attracted significant attention in recent years. These models require multiple sampling steps to reconstruct speech from noise, making inference speed a key challenge. Reducing the number of sampling steps can greatly improve inference efficiency. To this end, we introduce Fast F5-TTS, a training-free approach to accelerate the inference of flow-matching-based TTS models. By inspecting the sampling trajectory of F5-TTS, we identify redundant steps and propose Empirically Pruned Step Sampling (EPSS), a non-uniform time-step sampling strategy that effectively reduces the number of sampling steps. Our approach achieves a 7-step generation with an inference RTF of 0.030 on an NVIDIA RTX 3090 GPU, making it 4 times faster than the original F5-TTS while maintaining comparable performance. Furthermore, EPSS performs well on E2 TTS models, demonstrating its strong generalization ability.

Figures

Figures reproduced from arXiv: 2505.19931 by the authors.

Figure 1
Figure 1. PCA visualization of the 100-dimensional sampling trajectory during the inference process of Mel spectrogram. feature dimensions, providing insight into the generation dy￾namics. Using the 32-NFE trajectory as an example, two key obser￾vations emerge: Complex initial phase: The early trajectory exhibits signifi￾cant curvature, likely due to the model operating on near-noise inputs, leading to uncertainty in the flow… view at source ↗
Figure 2
Figure 2. Sampling trajectories for the 20th, 60th, and 80th dimensions of the 100-dimensional Mel spectrogram feature [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of EPSS sampling time-step distribution. 4. Experiments 4.1. Setup Baselines. We use officially open-sourced F5-TTS Base [11] as the baseline, which was trained on the Emilia dataset [25] for 1.2M updates to explore optimal settings for our EPSS method. The reproduced E2 TTS model from [11] is included for evalu￾ation and generalization study. Besides, we compared the per￾formance of ours with several o… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 21 canonical work pages

  1. [1]

    Introduction Recently, text-to-speech (TTS) has emerged as a promising approach for generating speech samples from a given input text while mimicking the voice characteristics of a reference speech. Existing methods have achieved significant break- throughs by scaling data and model size, enabling these sys- tems to generate speech of high fidelity and na...

    work page

  2. [2]

    Accelerating Flow-Matching-Based Text-to-Speech via Empirically Pruned Step Sampling

    Preliminaries 2.1. Formulation of flow matching LetR d be the data space andqbe the real data distribution. Flow matching learns a time-dependent vector fieldvt : [0,1]× Rd →R d, which generates a flowϕ t that transforms samples from a simple prior distributionp 0 to the target data distribution qviap 1 ≈q. This transformation is achieved by solving the O...

    work page Pith review arXiv 2025

  3. [3]

    Methods 3.1. Trajectories analysis of flow matching sampling Flow-matching-based generative models transform noise into real-world data through multiple neural network inference steps, with the time-steptprogressing from 0 to 1. Most flow matching models use a uniform time-step schedule for sam- pling, wheret k = k NFE , k= 0,1, ...,NFE. Although this str...

    work page

  4. [4]

    Setup Baselines

    Experiments 4.1. Setup Baselines. We use officially open-sourced F5-TTSBase[11] as the baseline, which was trained on the Emilia dataset [25] for 1.2M updates to explore optimal settings for our EPSS method. The reproduced E2 TTS model from [11] is included for evalu- ation and generalization study. Besides, we compared the per- formance of ours with seve...

    work page

  5. [5]

    Conclusion In this work, we introduce Fast F5-TTS, a simple and training- free approach to accelerate inference in flow-matching-based TTS models. By analyzing the sampling trajectory of F5-TTS, we propose Empirically Pruned Step Sampling (EPSS), a non- uniform time-step sampling strategy that prunes redundant sam- pling steps, to reduce the NFE while pre...

    work page

  6. [6]

    U23B2018 and No

    Acknowledgements This work was supported by the National Natural Science Foun- dation of China (No. U23B2018 and No. 62206171), Shang- hai Municipal Science and Technology Major Project under Grant 2021SHZDZX0102. The authors also acknowledge Bei- jng PARATERA Tech CO.,Ltd. for providing HPC resources that have contributed to the research results reported...

    work page

  7. [7]

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

    M. Le, A. Vyas, B. Shi, B. Karrer, L. Sari, R. Moritz, M. Williamson, V . Manohar, Y . Adi, J. Mahadeokaret al., “V oicebox: Text-guided multilingual universal speech generation at scale,”Advances in neural information processing systems, vol. 36, 2024

    2024

  8. [8]

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

    C. Wang, S. Chen, Y . Wu, Z.-H. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li, L. He, S. Zhao, and F. Wei, “Neural codec language models are zero-shot text to speech synthesizers,”IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 705–718, 2023

    work page 2023

Show all 38 references
  1. [9]

    BASE TTS: Lessons from building a billion-parameter text-to-speech model on 100k hours of data,

    M. Łajszczak, G. C ´ambara, Y . Li, F. Beyhan, A. van Korlaar, F. Yang, A. Joly, ´A. Mart ´ın-Cortinas, A. Abbas, A. Michalski et al., “BASE TTS: Lessons from building a billion-parameter text-to-speech model on 100k hours of data,”arXiv preprint arXiv:2402.08093, 2024

    2024 arXiv

  2. [10]

    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 (Volume 1: Long Papers), L.-W. Ku, A. Martins, and V...

    2024

  3. [11]

    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

    2024 arXiv

  4. [12]

    Autoregressive speech synthesis without vector quantization,

    L. Meng, L. Zhou, S. Liu, S. Chen, B. Han, S. Hu, Y . Liu, J. Li, S. Zhao, X. Wuet al., “Autoregressive speech synthesis without vector quantization,”arXiv preprint arXiv:2407.08551, 2024

    2024 arXiv

  5. [13]

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

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

    2024 arXiv

  6. [14]

    Flashspeech: Efficient zero-shot speech synthesis,

    Z. Ye, Z. Ju, H. Liu, X. Tan, J. Chen, Y . Lu, P. Sun, J. Pan, W. Bian, S. Heet al., “Flashspeech: Efficient zero-shot speech synthesis,” inProceedings of the 32nd ACM International Con- ference on Multimedia, 2024, pp. 6998–7007

    2024

  7. [15]

    NaturalSpeech 3: Zero-shot speech syn- thesis with factorized codec and diffusion models,

    Z. Ju, Y . Wang, K. Shen, X. Tan, D. Xin, D. Yang, E. Liu, Y . Leng, K. Song, S. Tanget al., “NaturalSpeech 3: Zero-shot speech syn- thesis with factorized codec and diffusion models,” inInterna- tional Conference on Machine Learning, 2024

    2024

  8. [16]

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

    S. E. Eskimez, X. Wang, M. Thakker, C. Li, C.-H. Tsai, Z. Xiao, H. Yang, Z. Zhu, M. Tang, X. Tanet al., “E2 TTS: Embarrassingly easy fully non-autoregressive zero-shot tts,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 682– 689

    2024

  9. [17]

    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

    2024 arXiv

  10. [18]

    F5-TTS: A fairytaler that fakes fluent and faith- ful 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 faith- ful speech with flow matching,”arXiv preprint arXiv:2410.06885, 2024

    2024 arXiv

  11. [19]

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

    K. Shen, Z. Ju, X. Tan, Y . Liu, Y . Leng, L. He, T. Qin, S. Zhao, and J. Bian, “Naturalspeech 2: Latent diffusion models are natu- ral and zero-shot speech and singing synthesizers,”International Conference on Learning Representations, 2023

    2023

  12. [20]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,”International Confer- ence on Learning Representations, 2022

    2022

  13. [21]

    Dmdspeech: Distilled diffu- sion model surpassing the teacher in zero-shot speech synthesis via direct metric optimization,

    Y . A. Li, R. Kumar, and Z. Jin, “Dmdspeech: Distilled diffu- sion model surpassing the teacher in zero-shot speech synthesis via direct metric optimization,”arXiv preprint arXiv:2410.11097, 2024

    2024 arXiv

  14. [22]

    Latent consistency models: Synthesizing high-resolution images with few-step infer- ence,

    S. Luo, Y . Tan, L. Huang, J. Li, and H. Zhao, “Latent consistency models: Synthesizing high-resolution images with few-step infer- ence,”arXiv preprint arXiv:2310.04378, 2023

    2023 arXiv

  15. [23]

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

    X. Liu, C. Gong, and qiang liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  16. [24]

    One-step diffusion with distribution matching distillation,

    T. Yin, M. Gharbi, R. Zhang, E. Shechtman, F. Durand, W. T. Freeman, and T. Park, “One-step diffusion with distribution matching distillation,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2024, pp. 6613–6623

    2024

  17. [25]

    Improved distribution matching distillation for fast image synthesis,

    T. Yin, M. Gharbi, T. Park, R. Zhang, E. Shechtman, F. Durand, and W. T. Freeman, “Improved distribution matching distillation for fast image synthesis,”Neural Information Processing Systems, 2024

    2024

  18. [26]

    V oiceflow: Efficient text-to-speech with rectified flow matching,

    Y . Guo, C. Du, Z. Ma, X. Chen, and K. Yu, “V oiceflow: Efficient text-to-speech with rectified flow matching,” inICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 121–11 125

    2024

  19. [27]

    All measurements are performed on single NVIDIA RTX 3090 GPU to ensure consistent hardware conditions

    as our test set, evaluating performance through four met- rics. All measurements are performed on single NVIDIA RTX 3090 GPU to ensure consistent hardware conditions. Word Error Rate (WER)measures the intelligibility of syn- thesized speech by comparing its transcription with ...

  20. [28]

    Autoregressive diffusion transformer for text-to-speech synthesis,

    Z. Liu, S. Wang, S. Inoue, Q. Bai, and H. Li, “Autoregressive diffusion transformer for text-to-speech synthesis,”arXiv preprint arXiv:2406.05551, 2024

    2024 arXiv

  21. [29]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  22. [30]

    Scalable diffusion models with transform- ers,

    W. Peebles and S. Xie, “Scalable diffusion models with transform- ers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4195–4205

    2023

  23. [31]

    Convnext v2: Co-designing and scaling convnets with masked autoencoders,

    S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, and S. Xie, “Convnext v2: Co-designing and scaling convnets with masked autoencoders,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2023, pp. 16 133–16 142

    2023

  24. [32]

    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

  25. [33]

    V ocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis,

    H. Siuzdak, “V ocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis,” International Conference on Learning Representations, 2023

    2023

  26. [34]

    LibriSpeech-PC: Benchmark for evaluation of punctuation and capitalization capabilities of end-to-end asr mod- els,

    A. Meister, M. Novikov, N. Karpov, E. Bakhturina, V . Lavrukhin, and B. Ginsburg, “LibriSpeech-PC: Benchmark for evaluation of punctuation and capitalization capabilities of end-to-end asr mod- els,” in2023 IEEE Automatic Speech Recognition and Under- standing Workshop (ASRU)....

    2023

  27. [35]

    Robust speech recognition via large-scale weak supervision,

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

    2023

  28. [36]

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

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

    2022 arXiv

  29. [37]

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

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  30. [38]

    UTMOS: Utokyo-sarulab system for voicemos challenge 2022,

    T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari, “UTMOS: Utokyo-sarulab system for voicemos challenge 2022,”Interspeech, 2022

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.