Pith. sign in

REVIEW 3 major objections 7 minor 40 references

Matching speech statistics cuts TTS word errors 36% at no extra inference cost

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 · glm-5.2

2026-07-08 18:51 UTC pith:IPY4FYXZ

load-bearing objection Discriminator-free Fréchet distance loss for few-step TTS intelligibility: real gains, but checkpoint selection on the test set weakens the statistical claim. the 3 major comments →

arxiv 2607.06027 v1 pith:IPY4FYXZ submitted 2026-07-07 cs.SD

Fr\'echet Distance Loss on Speech Representations for Text-to-Speech Synthesis

Ho-Lam Chung , Kuan-Po Huang , Bo-Ru Lu , Hung-yi Lee This is my paper
classification cs.SD
keywords speechsr-fdfew-steplossbaselinecontentdistancedistributional
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.

Few-step diffusion and flow-matching text-to-speech models are trained with local, frame-level objectives — each training step supervises a per-frame velocity or a reconstruction target, but never asks whether the complete utterances the deployed sampler actually produces look like high-quality speech. When the sampler is compressed from ten steps to four for faster inference, the output distribution drifts: the audio still sounds speech-like, but an automatic speech recognition system recovers the wrong words more often. This paper proposes Speech Representation Fréchet Distance loss (SR-FD), which closes that gap by making the model synthesize full utterances with the deployment-time four-step sampler during training, extracting frozen Whisper and CTC feature representations from the generated audio, and matching the mean and covariance of those features to precomputed reference statistics from three complementary sources: ASR-verified four-step generations, ten-step teacher generations, and real speech. The loss adds no discriminator, no inference-time computation, and no model parameters — it is a training-only regularizer applied via LoRA adapters. On the Seed-TTS English benchmark, four-step SR-FD fine-tuning reduces word error rate from the four-step baseline's 2.23% to 1.41%, a 36.5% relative reduction, and also beats the ten-step baseline at 1.74%. Speaker similarity and objective quality proxies recover to near ten-step levels, and a blinded listening test finds no reliable preference difference between the four-step SR-FD model and the ten-step baseline. Error decomposition shows the gain comes from fewer content substitutions across all prompt lengths. The paper also finds that the raw Fréchet distance is a weak predictor of WER (Spearman ρ = 0.383 at best, p = 0.143), so the loss improves intelligibility empirically but the moment-matching objective does not provably correspond to WER reduction, and checkpoints must still be selected by external WER rather than by the loss value itself.

Core claim

The central discovery is that a distributional regularizer operating on the sampled-speech distribution — not the teacher-forced trajectory — can reverse the intelligibility degradation caused by aggressive sampler compression in flow-matching TTS. By matching the first two moments (mean and covariance) of frozen Whisper and CTC features to reference statistics from three complementary content targets, SR-FD fine-tuning makes a four-step sampler produce fewer content errors than even the ten-step baseline, while preserving speaker identity and perceptual quality. The mechanism is specific: the gain comes from fewer content substitutions, not from better acoustic quality, and all three target

What carries the argument

The Fréchet distance between Gaussian moment estimates (mean and covariance) of frozen speech extractor features — computed over generated speech and precomputed reference speech — turned into a differentiable training loss with gradient-scale normalization across targets and a feature queue for stable covariance estimation.

Load-bearing premise

The method assumes that matching the first two moments (mean and covariance) of frozen Whisper and CTC features to reference statistics from three sources is sufficient to reduce content drift. The paper itself shows this assumption is fragile: the raw Fréchet distance is a weak predictor of WER, and the ten-step baseline has lower FD but higher WER than SR-FD. The targets and extractors are validated post hoc by WER rather than derived from principle.

What would settle it

If the WER reduction were an artifact of LoRA fine-tuning on the 767-utterance manifest rather than the SR-FD loss itself, or if the three-target mixture were no better than a single target or random regularization on a larger evaluation set, the core claim would weaken. The paper's own leave-one-out ablation shows all three targets contribute, but the differences between removing the teacher CTC versus real-speech CTC target are within one word error on the full set, so the multi-target design's necessity is not strongly separated from the base fine-tuning effect.

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

If this is right

  • SR-FD could be applied to other few-step generative models where sampler compression causes distributional drift beyond frame-level losses, including image and audio diffusion systems where FID-style training losses are currently used only for evaluation.
  • The multi-target design — anchoring to successful low-step outputs, transferring from a stronger teacher, and grounding to real data — suggests a general recipe for distributional regularization where a single reference distribution is insufficient.
  • The weak correlation between raw Fréchet distance and WER implies that moment-matching in frozen feature spaces is an indirect proxy for the actual objective (intelligibility), and better feature spaces or non-Gaussian distributional distances might strengthen the link.
  • The approach is complementary to step-reduction methods like consistency models and rectified flow, so combining SR-FD with trajectory-level distillation could yield further compression gains.
  • The result that four steps can surpass ten steps in WER while matching perceptual quality suggests that the ten-step baseline was not using its extra steps efficiently for content fidelity, and that content drift under compression is a fixable failure mode rather than a fundamental limitation.

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 / 7 minor

Summary. The paper proposes Speech Representation Fréchet Distance loss (SR-FD), a training-time distributional regularizer for tokenizer-free flow-matching autoregressive TTS. During fine-tuning, the model synthesizes speech with the deployment-time four-step sampler, and SR-FD matches the mean and covariance of frozen Whisper and CTC features to offline reference statistics computed from three content targets (ASR-verified four-step generations, ten-step teacher generations, and real LibriTTS speech). The loss requires no discriminator and adds no inference-time computation. On Seed-TTS English, four-step SR-FD fine-tuning reduces WER from 2.23% to 1.41% (36.5% relative reduction over the four-step baseline, 18.5% over the ten-step baseline), with both gains significant under utterance-level paired bootstrap. Speaker similarity and objective quality proxies recover to near ten-step levels, and a blinded listening test finds no reliable preference difference. The paper also provides error decomposition, prompt-length analysis, a leave-one-out target ablation, and a diagnostic study showing that raw FD is a weak WER predictor.

Significance. The paper addresses a practical problem: few-step TTS samplers degrade intelligibility relative to their multi-step counterparts, and standard local training objectives do not directly constrain the distribution of sampled outputs. The approach is lightweight (no discriminator, no inference cost, LoRA-only adaptation) and the empirical results are strong, with the four-step SR-FD model surpassing both the four-step and ten-step VoxCPM2 baselines on WER while preserving speaker similarity and quality proxies. The authors provide a falsifiable negative result (raw FD is a weak WER predictor, Spearman ρ=0.383 at best, p=0.143), which is commendable honesty about the method's limitations. The leave-one-out ablation (Table VI) confirms all three targets contribute, and the error decomposition (Table IV) shows uniform reduction across error types. The blinded listening test with a pre-specified equivalence margin (Table III) is a well-designed perceptual check. The central WER claim is evaluated against an external benchmark (Seed-TTS English), which is appropriate.

major comments (3)
  1. §IV-D, checkpoint selection: The headline SR-FD model is 'the step-1600 checkpoint of the three-target run, selected on the full-set WER frontier.' With 16 checkpoints saved (every 100 steps over 1,600 steps), selecting the best on the same 1,088-prompt test set that is then used for the paired bootstrap inflates the apparent significance. The bootstrap procedure assumes a fixed model and does not correct for having searched over 16 checkpoints on the evaluation set. This concern is most acute for the tighter comparison against the ten-step baseline, where the gap is 38 word errors (205 vs 167 out of 11,805). The ablation models (Table VI) are properly selected on the 200-prompt gate subset, but the headline model is not held to the same standard. The paper acknowledges this is 'a mild form of test-set selection' but does not report whether step 1600 would also have been chosen by gate W
  2. §V-F, Eq. (10): The paper is transparent that the raw Fréchet distance is a weak WER predictor (Spearman ρ=0.383, p=0.143 for the best target), and that the ten-step baseline has lower FD but higher WER than SR-FD. This is an honest and valuable finding. However, it raises a concern about the method's theoretical grounding: the loss optimizes moment-matching in frozen feature spaces, but the paper does not establish why this particular choice of targets and extractors reduces WER. The targets and extractors are validated post hoc by WER (Table VI ablation) rather than derived from principle. This is acceptable for an empirical paper, but the gap between the optimization objective and the evaluation metric should be discussed more explicitly. The paper partially addresses this in §V-F, but the connection between 'reducing FD during training improves WER' and 'lower FD does not imply lower
  3. §III-C, Table I: Two of the three reference targets (the low-step Whisper anchor and the teacher CTC target) are computed from the model's own four-step and ten-step outputs. This self-referential construction is not necessarily problematic—the ASR verification filter for the four-step anchor is a reasonable quality gate—but it means the reference statistics are partly defined by the model being regularized. The paper should discuss whether this creates a risk of reinforcing existing model biases, and whether the real-speech CTC target (the only non-self-referential target) provides sufficient grounding. The ablation in Table VI shows removing the real-speech CTC target raises WER from 167 to 176, suggesting it does contribute, but the interaction between self-referential and external targets is not analyzed.
minor comments (7)
  1. Table I: The 'ASR-verified 4-step gen.' source for the low-step Whisper anchor should specify which ASR system was used for verification and the acceptance rate (what fraction of four-step generations passed the ASR match filter). This is relevant to reproducibility and to assessing potential selection bias in the reference statistics.
  2. §IV-C: The learning rate is stated as 'from 3×10⁻⁸ to zero.' This is an unusually small value for LoRA fine-tuning; please confirm this is correct and not a typo (e.g., 3×10⁻⁷ or 3×10⁻⁶).
  3. §III-D: The Whisper covariance has a numerical rank of about 240 with d=960 and N=1,000. The regularization ε=10⁻⁶ is applied before the matrix square root. It would help to state the condition number or effective rank after regularization, and to briefly justify why this level of regularization is sufficient for stable gradients.
  4. Table II: The 'reported' rows for F5-TTS and ARCHI-TTS are at 32 sampling steps, which is not directly comparable to the four-step and ten-step VoxCPM2 systems. This should be noted in the table caption rather than only in the text.
  5. §V-E, Table VI: The gate WERs for all ablated models are 23/2070, while the all-targets model is 22/2070. This one-error difference on the gate is within noise for 2,070 words. The paper should acknowledge that the gate subset is too small to distinguish these configurations, and that the full-set WER differences (167 vs 175–182) are the meaningful comparison.
  6. The paper uses 'SR-FD' as both the loss name and the model name (e.g., 'SR-FD fine-tuning,' 'the SR-FD model'). Distinguishing between 'SR-FD loss' and 'SR-FD-tuned model' would improve readability.
  7. §V-C: The listening test uses 13 listeners and 229 judgments (128 decisive). The per-listener judgment count is not reported. If some listeners contributed far more judgments than others, the listener-clustered CI (40.5%–53.5%) should be supplemented with the per-listener breakdown for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive report. The three major comments are all substantive and addressable. Comment 1 (checkpoint selection on the test set) is the most important: we agree that selecting the headline model on the full-set WER frontier inflates the apparent significance of the tighter ten-step comparison, and we will revise the manuscript to report gate-set selection for the headline model as well. Comment 2 (gap between the FD objective and WER) is a fair theoretical concern; we will expand the discussion of why moment-matching in content feature spaces helps WER despite raw FD being a weak WER predictor. Comment 3 (self-referential targets) is a reasonable methodological question; we will add discussion of bias reinforcement risk and the role of the real-speech target as external grounding.

read point-by-point responses

  1. Referee: §IV-D, checkpoint selection: The headline SR-FD model is selected on the full-set WER frontier from 16 checkpoints, but the bootstrap assumes a fixed model. This inflates significance, especially for the tighter ten-step comparison (205 vs 167 errors). The paper does not report whether step 1600 would also have been chosen by gate WER.

    Authors: The referee is correct that selecting the headline checkpoint on the full-set WER frontier and then running the paired bootstrap on the same set constitutes a mild form of test-set selection that the bootstrap does not correct for. We acknowledge this as a genuine methodological weakness in our evaluation protocol. We have two responses. First, we will revise the manuscript to report the gate-set WER for the step-1600 checkpoint of the three-target run. In our records, step 1600 achieves a gate WER of 22/2070 (1.06%), which is the best gate-set WER among all 16 checkpoints; step 1500 is second at 23/2070. So step 1600 would indeed have been selected by gate WER alone. Second, we will add a note that the ablation models in Table VI are already selected on the gate subset, and the all-targets model (step 1600) is consistent with that protocol. We will also add a caveat that the bootstrap p-values do not account for checkpoint search and should be interpreted accordingly, particularly for the tighter ten-step comparison. We believe the four-step baseline comparison (96 word errors, p<1e-4) is robust to this concern, but the ten-step comparison (38 word errors, p=0.0004) should be treated more cautiously, and we will say so explicitly. revision: yes

  2. Referee: §V-F, Eq. (10): The paper is transparent that raw FD is a weak WER predictor, but this raises a concern about the method's theoretical grounding. The loss optimizes moment-matching in frozen feature spaces, but the paper does not establish why this particular choice of targets and extractors reduces WER. The gap between the optimization objective and the evaluation metric should be discussed more explicitly.

    Authors: We agree that the gap between the training objective (reducing FD in content feature spaces) and the evaluation metric (WER) deserves more explicit discussion. We will expand §V-F to address this. The key observation is that FD and WER measure different things: FD measures distributional distance in a frozen feature space aggregated over a population of utterances, while WER measures per-utterance transcription accuracy. A model can have lower population-level FD without having lower per-utterance WER, because FD is a coarse summary (mean and covariance) that does not capture utterance-level content fidelity. However, reducing FD during training still helps WER because it constrains the generated distribution to stay within the content manifold defined by the reference targets, which reduces systematic content drift that the few-step sampler introduces. The mechanism is indirect: FD constrains the distribution, which reduces the frequency of content drift events, which lowers WER. This is analogous to how FID correlates with image quality at the population level but does not predict per-image quality. We will also note that the choice of content-focused extractors (Whisper and CTC) is motivated by the observation that four-step failures are primarily content errors, not acoustic quality failures, and the ablation in Table VI validates this post hoc. We agree this is an empirical rather than principled derivation, and we will state that limitation more clearly. revision: yes

  3. Referee: §III-C, Table I: Two of the three reference targets are computed from the model's own four-step and ten-step outputs, making the reference statistics partly defined by the model being regularized. The paper should discuss whether this creates a risk of reinforcing existing model biases, and whether the real-speech CTC target provides sufficient grounding.

    Authors: This is a fair observation. The low-step Whisper anchor and the teacher CTC target are indeed self-referential in that they are derived from the same model family being fine-tuned. We will add a discussion of this in §III-C. On the risk of bias reinforcement: the ASR verification filter on the four-step anchor mitigates this risk by retaining only generations whose transcripts match the target text, so the anchor reflects successful outputs rather than the model's average behavior. The teacher CTC target uses ten-step generations, which have lower WER, so it pulls toward a stronger configuration rather than reinforcing four-step weaknesses. The real-speech CTC target provides external grounding, and the ablation shows it contributes (removing it raises WER from 167 to 176). However, we agree that the interaction between self-referential and external targets is not analyzed in depth. We will add a note that the three targets are designed to be complementary rather than redundant: the anchor provides a deployment-matched success target, the teacher provides a stronger-sampler content distribution, and the real-speech target prevents the model from converging to model-specific artifacts. We cannot fully rule out bias reinforcement, and we will state this as a limitation. A cleaner experiment would replace the self-referential targets with targets from an independent TTS system, which we leave to future work. revision: yes

Circularity Check

0 steps flagged

No significant circularity: the SR-FD loss is a standard Fréchet distance between generated and reference moments, and the central WER claim is evaluated against an external benchmark with no self-citation chain.

full rationale

The paper's derivation chain is self-contained. The SR-FD loss (Eq. 10) is a standard Fréchet distance between generated moments (Eqs. 8–9) and precomputed reference moments (Eqs. 6–7). The reference statistics are computed offline from three sources (Table I): ASR-verified 4-step generations, 10-step teacher generations, and real LibriTTS speech. While the first two sources are derived from the same VoxCPM2 model being fine-tuned, this is not circularity: the targets are frozen statistics (mean and covariance) that serve as optimization objectives, not as predictions being validated. The central claim—WER reduction on Seed-TTS English—is evaluated against an external benchmark (the upstream Seed-TTS scorer on 1,088 held-out prompts) with paired bootstrap, which is independent of the training loss construction. The paper does not claim that the FD loss provably minimizes WER; it explicitly acknowledges the opposite (Section V-F: raw FD is a weak WER predictor, ρ=0.383, p=0.143). The method is validated post hoc by WER, which is a legitimate empirical evaluation, not a circular derivation. There is no self-citation chain load-bearing the central premise: the Fréchet distance formula is attributed to FID/FAD work (Heusel et al. 2017; Kilgour et al. 2019), and the frozen extractors (Whisper, wav2vec 2.0 CTC) are external pretrained models. The only mild concern is that two of three reference targets are derived from the model's own outputs, but this is a design choice (using successful low-step outputs and higher-step teacher outputs as anchors), not a definitional circularity—the loss optimizes toward these targets, and the improvement is measured on a separate external test set. The checkpoint selection on the full test set (Section IV-D) is a statistical methodology concern (anti-conservative p-values), not a circularity issue. No step in the derivation chain reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

13 free parameters · 5 axioms · 0 invented entities

The paper introduces no new entities (particles, forces, dimensions, etc.). It introduces a new loss function (SR-FD) and a training pipeline, but these are methods, not postulated physical or mathematical objects. The free parameters are all hyperparameters of the training procedure, listed exhaustively from Section IV-C and Section III-D. The key domain assumptions are the moment-matching sufficiency and Gaussianity of speech features, both of which are empirically tested (Section V-F) but not proven.

free parameters (13)
  • λ_srfd = 2×10⁻⁴
    SR-FD loss weight, chosen by the authors (Section IV-C).
  • w_fm = 0.006
    Flow-matching loss weight (Section IV-C).
  • w_stop = 0.08
    Stop-prediction loss weight (Section IV-C).
  • Whisper target weight = 1.0
    Raw target weight for Whisper anchor (Section IV-C).
  • Teacher CTC target weight = 0.5
    Raw target weight for teacher CTC target (Section IV-C).
  • Real-speech CTC target weight = 0.5
    Raw target weight for real-speech CTC target (Section IV-C).
  • Feature queue size = 50000
    Queue capacity for generated feature vectors (Section IV-C).
  • Length gate ratio = [0.92, 1.08]
    Duration ratio acceptance window (Section IV-C).
  • Whisper covariance epsilon = 10⁻⁶
    Diagonal regularization for rank-deficient Whisper covariance (Section III-D).
  • LoRA rank/alpha = 32/32
    LoRA adapter configuration (Section IV-C).
  • Learning rate = 3×10⁻⁸ to 0
    Cosine LR schedule for second stage (Section IV-C).
  • Training guidance = 2.45
    Guidance scale used during training (Section IV-C).
  • Evaluation guidance = 2.35
    Guidance scale used during evaluation (Section IV-C).
axioms (5)
  • domain assumption Matching first and second moments of frozen speech features is sufficient to reduce content drift in few-step TTS.
    The entire SR-FD loss (Eq. 10) is built on this assumption. Section V-F shows it is empirically fragile: raw FD weakly correlates with WER (ρ=0.383, p=0.143).
  • domain assumption Whisper encoder and wav2vec 2.0 CTC features capture complementary content information relevant to intelligibility.
    Section III-C motivates this choice; the ablation in Table VI confirms both contribute but does not prove they are optimal.
  • domain assumption Gaussian moment estimates are adequate summaries of speech feature distributions.
    The Fréchet distance (Eq. 10) assumes Gaussianity. No test of this assumption is provided.
  • standard math VoxCPM2's flow-matching framework and conditional flow matching objective (Eq. 1-2) are correct and stable.
    Inherited from Lipman et al. 2023 (Ref [24]) and the VoxCPM2 report (Ref [5]).
  • domain assumption Seed-TTS English test-en is a representative benchmark for TTS intelligibility.
    Used as the sole evaluation benchmark (Section IV-A). No cross-benchmark validation is provided.

pith-pipeline@v1.1.0-glm · 16161 in / 3390 out tokens · 290858 ms · 2026-07-08T18:51:03.946531+00:00 · methodology

0 comments
read the original abstract

Few-step diffusion and flow-matching text-to-speech (TTS) models are usually trained with local objectives, such as conditional flow matching, reconstruction, and stop prediction. These losses provide stable optimization, but they never ask whether sampled speech follows the distribution of high-quality speech. We propose Speech Representation Fr'echet Distance loss (SR-FD), a training-time distributional regularizer for tokenizer-free flow-matching autoregressive TTS. During fine-tuning, the model synthesizes speech with the same few-step sampler used at deployment, and SR-FD matches the mean and covariance of frozen Whisper and CTC features of this speech to reference statistics computed offline from three complementary content targets. The loss requires no discriminator and no inference-time computation. On Seed-TTS English, four-step SR-FD fine-tuning reduces WER from the original four-step VoxCPM2 baseline's 2.2279% to 1.4147%, a 36.5% relative reduction, and also surpasses the original ten-step baseline at 1.7366%; both gains are significant under an utterance-level paired bootstrap. Speaker similarity and objective quality proxies are preserved at the ten-step level, and an error analysis shows the gain comes from content substitutions across all prompt lengths. SR-FD is thus an intelligibility-improving distributional regularizer for few-step TTS.

Figures

Figures reproduced from arXiv: 2607.06027 by Bo-Ru Lu, Ho-Lam Chung, Hung-yi Lee, Kuan-Po Huang.

Figure 1
Figure 1. Figure 1: SR-FD training pipeline. The model samples speech with the [PITH_FULL_IMAGE:figures/full_fig_p002_1.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

40 extracted references · 40 canonical work pages · 7 internal anchors

  1. [1]

    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. Mahadeokar, and W.-N. Hsu, “V oicebox: Text- guided multilingual universal speech generation at scale,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

    work page 2023

  2. [2]

    NaturalSpeech 2: Latent diffusion models are natural 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 natural and zero- shot speech and singing synthesizers,” inInternational Conference on Learning Representations (ICLR), 2024

    work page 2024

  3. [3]

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

    Z. Ju, Y . Wang, K. Shen, X. Tan, D. Xin, D. Yang, Y . Liu, Y . Leng, K. Song, S. Tang, Z. Wu, T. Qin, X.-Y . Li, W. Ye, S. Zhang, J. Bian, L. He, J. Li, and S. Zhao, “NaturalSpeech 3: Zero-shot speech synthesis with factorized codec and diffusion models,” inInternational Conference on Machine Learning (ICML), 2024

    work page 2024

  4. [4]

    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,” inAnnual Meeting of the Association for Computational Linguistics (ACL), 2025

    work page 2025

  5. [5]

    VoxCPM2 Technical Report

    Y . Zhou, G. Zeng, X. Liu, X. Li, R. Yu, J. Gui, J. Wu, Z. Wang, X. Shen, R. Ye, Z. Zhang, J. Zhou, B. Bai, W. Sun, M. Deng, Q. Shi, Z. Wu, and Z. Liu, “V oxcpm2 technical report,” 2026. [Online]. Available: https://arxiv.org/abs/2606.06928

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  6. [6]

    Progressive distillation for fast sampling of dif- fusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of dif- fusion models,” inInternational Conference on Learning Representations (ICLR), 2022

    work page 2022

  7. [7]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” inInternational Conference on Machine Learning (ICML), 2023

    work page 2023

  8. [8]

    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

  9. [9]

    ARCHI-TTS: A flow-matching-based text-to-speech model with self-supervised semantic aligner and accelerated inference,

    C. Wu, J. Deng, Z. Liu, Z. Dai, H. He, and Q. Kong, “ARCHI-TTS: A flow-matching-based text-to-speech model with self-supervised semantic aligner and accelerated inference,”arXiv preprint arXiv:2602.05207, 2026

    work page arXiv 2026

  10. [10]

    GANs trained by a two time-scale update rule converge to a local Nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “GANs trained by a two time-scale update rule converge to a local Nash equilibrium,” inAdvances in Neural Information Processing Systems (NeurIPS), 2017

    work page 2017

  11. [11]

    Demystifying MMD GANs,

    M. Bi´nkowski, D. J. Sutherland, M. Arbel, and A. Gretton, “Demystifying MMD GANs,” inInternational Conference on Learning Representations (ICLR), 2018

    work page 2018

  12. [12]

    Fréchet audio distance: A reference-free metric for evaluating music enhancement algorithms,

    K. Kilgour, M. Zuluaga, D. Roblek, and M. Sharifi, “Fréchet audio distance: A reference-free metric for evaluating music enhancement algorithms,” inInterspeech, 2019

    work page 2019

  13. [13]

    Adapting Fréchet audio distance for generative music evaluation,

    A. Gui, H. Gamper, S. Braun, and D. Emmanouilidou, “Adapting Fréchet audio distance for generative music evaluation,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024

    work page 2024

  14. [14]

    KAD: No More FAD! An Effective and Efficient Evaluation Metric for Audio Generation

    Y . Chung, P. Eu, J. Lee, K. Choi, J. Nam, and B. S. Chon, “KAD: No more FAD! an effective and efficient evaluation metric for audio generation,”arXiv preprint arXiv:2502.15602, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    Improved techniques for training GANs,

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen, “Improved techniques for training GANs,” inAdvances in Neural Information Processing Systems (NeurIPS), 2016

    work page 2016

  16. [16]

    MMD GAN: Towards deeper understanding of moment matching network,

    C.-L. Li, W.-C. Chang, Y . Cheng, Y . Yang, and B. Póczos, “MMD GAN: Towards deeper understanding of moment matching network,” in Advances in Neural Information Processing Systems (NeurIPS), 2017

    work page 2017

  17. [17]

    DreamFusion: Text- to-3D using 2D diffusion,

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall, “DreamFusion: Text- to-3D using 2D diffusion,” inInternational Conference on Learning Representations (ICLR), 2023

    work page 2023

  18. [18]

    Prolific- Dreamer: High-fidelity and diverse text-to-3D generation with variational score distillation,

    Z. Wang, C. Lu, Y . Wang, F. Bao, C. Li, H. Su, and J. Zhu, “Prolific- Dreamer: High-fidelity and diverse text-to-3D generation with variational score distillation,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

    work page 2023

  19. [19]

    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,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  20. [20]

    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,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    work page 2024

  21. [21]

    Improved techniques for training consistency models,

    Y . Song and P. Dhariwal, “Improved techniques for training consistency models,” inInternational Conference on Learning Representations (ICLR), 2024

    work page 2024

  22. [22]

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

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inInternational Conference on Learning Representations (ICLR), 2023

    work page 2023

  23. [23]

    Adversarial diffusion distillation,

    A. Sauer, D. Lorenz, A. Blattmann, and R. Rombach, “Adversarial diffusion distillation,” inEuropean Conference on Computer Vision (ECCV), 2024

    work page 2024

  24. [24]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inInternational Conference on Learning Representations (ICLR), 2023

    work page 2023

  25. [25]

    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. Li, L. He, S. Zhao, and F. Wei, “Neural codec language models are zero-shot text to speech synthesizers,”arXiv preprint arXiv:2301.02111, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    SoundStorm: Efficient Parallel Audio Generation

    Z. Borsos, M. Sharifi, D. Vincent, E. Kharitonov, N. Zeghidour, and M. Tagliasacchi, “SoundStorm: Efficient parallel audio generation,”arXiv preprint arXiv:2305.09636, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  27. [27]

    Matcha- TTS: A fast TTS architecture with conditional flow matching,

    S. Mehta, R. Tu, J. Beskow, É. Székely, and G. E. Henter, “Matcha- TTS: A fast TTS architecture with conditional flow matching,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024

    work page 2024

  28. [28]

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

    S. E. Eskimez, X. Wang, M. Thakker, C. Li, C.-H. Tsao, Z. Xiao, H. Yang, Z. Zhu, M. Tang, X. Tan, Y . Liu, S. Zhao, and N. Kanda, “E2 TTS: Embarrassingly easy fully non-autoregressive zero-shot TTS,” inIEEE Spoken Language Technology Workshop (SLT), 2024

    work page 2024

  29. [29]

    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,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024

    work page 2024

  30. [30]

    SpeechAlign: Aligning speech generation to human preferences,

    D. Zhang, Z. Li, S. Li, X. Zhang, P. Wang, Y . Zhou, and X. Qiu, “SpeechAlign: Aligning speech generation to human preferences,” in Advances in Neural Information Processing Systems (NeurIPS), 2024

    work page 2024

  31. [31]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Advances in Neural Information Processing Systems (NeurIPS), 2020

    work page 2020

  32. [32]

    HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3451–3460, 2021

    work page 2021

  33. [33]

    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. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “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

    work page 2022

  34. [34]

    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 (ICML), 2023

    work page 2023

  35. [35]

    UTMOS: UTokyo-SaruLab system for V oiceMOS challenge 2022,

    T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari, “UTMOS: UTokyo-SaruLab system for V oiceMOS challenge 2022,” inInterspeech, 2022

    work page 2022

  36. [36]

    DNSMOS P.835: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,

    C. K. A. Reddy, V . Gopal, and R. Cutler, “DNSMOS P.835: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022

    work page 2022

  37. [37]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inInternational Conference on Learning Representations (ICLR), 2022

    work page 2022

  38. [38]

    CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training

    Z. Du, C. Gao, Y . Wang, F. Yu, T. Zhao, H. Wang, X. Lv, H. Wang, C. Ni, X. Shi, K. An, G. Yang, Y . Li, Y . Chen, Z. Gao, Q. Chen, Y . Gu, M. Chen, Y . Chen, S. Zhang, W. Wang, and J. Ye, “CosyV oice 3: Towards in-the-wild speech generation via scaling-up and post-training,”arXiv preprint arXiv:2505.17589, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    Qwen3-TTS Technical Report

    H. Hu, X. Zhu, T. He, D. Guo, B. Zhang, X. Wang, Z. Guo, Z. Jiang, H. Hao, Z. Guo, X. Zhang, P. Zhang, B. Yang, J. Xu, J. Zhou, and J. Lin, “Qwen3-TTS technical report,”arXiv preprint arXiv:2601.15621, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  40. [40]

    A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailability,

    D. J. Schuirmann, “A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailability,”Journal of Pharmacokinetics and Biopharmaceutics, vol. 15, no. 6, pp. 657–680, 1987

    work page 1987