Pith. sign in

REVIEW 2 major objections 6 minor 31 references

By moving flow-matching speech generation into a 25 Hz latent space, ZipL-Dialog cuts peak GPU memory by 11.22x and speeds up long-form dialog synthesis by 2.23x while keeping naturalness.

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

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

T0 review · deepseek-v4-flash

2026-08-02 06:27 UTC pith:TX6SNWHM

load-bearing objection A credible efficiency win (latent CFM at 25 Hz) with a real but unproven quality claim on long-form dialog; worth engaging, not worth trusting yet. the 2 major comments →

arxiv 2607.12496 v2 pith:TX6SNWHM submitted 2026-07-14 eess.AS

ZipL-Dialog: Memory-Efficient Long-Form Spoken Dialog Synthesis via Latent Flow Matching

Jihwan Kim , Nam Soo Kim This is my paper
classification eess.AS
keywords text-to-speechdialogue synthesisflow matchinglatent spacememory efficiencyconditional flow matchingmel autoencoderzero-shot TTS
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.

Long-form spoken dialog synthesis with flow-matching models usually runs on dense 100 Hz mel-spectrograms, so memory and compute balloon with duration, often forcing chunked generation. This paper tries to establish that the same generation can be done in a 4x time-compressed 25 Hz latent space without losing competitive naturalness. The core claim is that a deterministic mel autoencoder with auxiliary mel-domain supervision, plus a rebalanced downsampling schedule, preserves enough phonetic and speaker detail under compression. If true, multi-minute dialogs can be synthesized in a single pass using a fraction of the GPU memory (up to 11.22x less peak) and about twice as fast, making long-context conversational TTS practical on modest hardware.

Core claim

The paper's central claim is that moving conditional flow matching for zero-shot dialog TTS into a 4x time-compressed continuous latent domain removes the sequence-length memory bottleneck of frame-level generation while keeping perceived naturalness. The model encodes mel-spectrograms into a 25 Hz deterministic latent space, trains a masked velocity-matching objective there, and adds an auxiliary mel-reconstruction loss so the decoder keeps local acoustic detail. On two dialog benchmarks, this reduces maximum peak GPU memory by 11.22x and speeds inference by 2.23x relative to the frame-level flow-matching baseline, with the best or tied-best UTMOS scores despite modestly higher WER and lowe

What carries the argument

The key mechanism is latent conditional flow matching: a frozen deterministic mel autoencoder compresses 100 Hz mel-spectrograms 4x into a 25 Hz latent sequence; the flow decoder predicts the velocity field on the masked target region in that latent space; and an auxiliary mel-domain reconstruction loss, weighted at 0.5, pulls the decoded latents back toward acoustic fidelity. The ZipFormer backbone uses a deliberately mild [1,1,2,1,1] downsampling schedule, since the default schedule designed for frame-level input over-compresses short phonemes once the input is already compressed.

Load-bearing premise

The frozen autoencoder was trained only on 3-second audio crops yet must reconstruct phonetics and speaker identity for dialogs up to 179 seconds at 4x temporal compression; if the compression degrades on long horizons, the perceptual-quality claim fails.

What would settle it

Run the frozen 4x autoencoder on held-out dialogs of 30, 60, 120, and 180 seconds and measure reconstruction WER and speaker similarity; if intelligibility or speaker similarity degrades sharply with duration, the long-horizon compression assumption fails.

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

If this is right

  • Minute-scale dialogs can be synthesized in a single pass on a single 40 GB GPU using about 1 GB average and 3.23 GB peak memory, versus 4-6 GB for the frame-level baseline.
  • The 2.23x inference speedup and real-time factor around 0.05 bring near-real-time long-form dialog synthesis within reach for interactive applications.
  • Because the flow model operates at 25 Hz, sequence-length-related memory grows far more slowly with duration, extending the practical horizon for conversational context modeling.
  • The quality trade-off is localized: WER rises modestly (5.20% vs 4.23% on one benchmark), so future improvements to latent reconstruction could close the gap while keeping the efficiency gains.

Where Pith is reading between the lines

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

  • The autoencoder was trained on 3-second crops, so the paper does not directly demonstrate reconstruction quality at the 179-second end of the test range; a natural next experiment is to plot reconstruction WER and speaker similarity against utterance position and duration.
  • All memory numbers are reported at batch size 1; the relative advantage of latent-space generation could be larger during training with long sequences, or smaller at high batch sizes where activation memory is amortized — worth benchmarking.
  • The drop in cpSIM (speaker-similarity) suggests the compressed latent blurs speaker-specific timbre; adding a speaker-identity or perceptual loss to the autoencoder decoder would be a direct testable extension.
  • The same 25 Hz latent recipe could generalize to other long-form audio generation tasks such as audiobooks, podcasts, or simultaneous interpretation, where chunked generation is currently the workaround.

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

2 major / 6 minor

Summary. The paper proposes ZipL-Dialog, a non-autoregressive latent conditional flow-matching TTS system for multi-minute spoken dialog. A deterministic mel autoencoder compresses 100 Hz mel frames by a factor of 4 into a 25 Hz latent sequence; a ZipFormer-based flow decoder then performs masked conditional flow matching in this latent space, with an auxiliary mel-domain reconstruction loss and a redesigned hierarchical downsampling schedule. On the CoVoMix2 and OpenDialog test sets, the system is reported to reduce maximum peak GPU memory by up to 11.22x and inference time by up to 2.23x relative to ZipVoice-Dialog, while achieving tied or best UTMOS. Quality is also assessed via WER and cpSIM, where ZipL-Dialog is somewhat worse than the frame-level baseline. Ablations in a single-speaker TTS recipe compare latent types, auxiliary loss, and downsampling schedules.

Significance. If the efficiency numbers are reproducible, the memory reduction is a practically meaningful advance: it makes single-pass synthesis of minute-scale dialog feasible on a single 40 GB GPU, addressing a real bottleneck of frame-level CFM. The paper provides a direct comparison with the closest baseline, a clear efficiency table, and ablations that support the chosen latent formulation and downsampling schedule. However, the central quality claim—that perceptual naturalness is maintained—is currently supported only by a single learned metric (UTMOS) with no error bars or human listening, and it is not demonstrated that the frozen autoencoder trained on 3-second crops preserves fidelity at 179-second test items. These gaps need to be closed before the conclusion can be accepted.

major comments (2)
  1. [§3.1, §4.2, Tables 1–2] The paper claims to maintain perceptual naturalness specifically for multi-minute dialog, but the frozen deterministic autoencoder is trained only on 3 s random crops (Section 4.2), i.e., 75 latent tokens at 25 Hz, and is then applied to target sequences up to 178.9 s (~4,473 latent tokens). The manuscript reports no AE reconstruction metrics versus duration, no positional embedding range analysis, and no evaluation of whether the encoder/decoder preserves phonetics and speaker identity at long horizons. The observed WER degradation on CoVoMix2 (5.20% vs 3.42%) and OpenDialog (5.36% vs 3.55%) in Table 2 is consistent with AE quality collapse on long sequences, though it could also come from the flow decoder. Please provide duration-binned quality metrics (e.g., WER/cpSIM/UTMOS by target length) and AE-only reconstruction metrics (mel distance and WER of reconstructed ground-truth audio)
  2. [§4.6, Table 2] The quality conclusion is under-powered. ZipL-Dialog is worse than ZipVoice-Dialog on WER and cpSIM on both test sets, so the 'maintaining perceptual naturalness' claim relies on UTMOS being tied or highest (3.523 vs 3.477 on CoVoMix2; 3.198 vs 3.089 on OpenDialog). UTMOS is a single non-linear predictor and the differences are small; no confidence intervals, significance tests, or human listening results are reported. Please add error bars and significance testing, and preferably a human MOS or AB preference test on long dialog samples. Without this, the central quality claim is not established.
minor comments (6)
  1. [§4.5, Ref. [17]] The WER metric says 'WhisperD [17]' but reference [17] is a blog post about Parakeet. Please clarify which ASR model was actually used and its configuration.
  2. [§4.4, Table 1] Please report whether the peak memory numbers include CUDA context and whether the VibeVoice baseline uses a different vocoder or sampling scheme; this helps interpret the efficiency comparison.
  3. [§4.7, Table 3] The no-downsampling and default schedules yield WER 51.96 and 27.43, respectively, which are far outside normal TTS intelligibility ranges. Please verify these settings did not suffer from training instability or ASR collapse and describe what the generated audio sounded like.
  4. [§4.2] The AE is described as 'deterministic'; please specify the exact training loss terms for the AE (e.g., mel reconstruction loss, any adversarial or perceptual losses) and give enough architectural detail for reproduction.
  5. [§3.2, §4.5] Inference uses 16-step Euler with classifier-free guidance; please specify the guidance scale and whether it was tuned separately for the baseline and proposed model.
  6. [§4.3] The OpenDialog test set construction should be described or cited; currently only 'used in ZipVoice-Dialog' is given.

Circularity Check

0 steps flagged

No significant circularity: efficiency and quality claims are measured against independent baselines and external test sets, with no fitted parameter renamed as a prediction.

full rationale

ZipL-Dialog's central claims are empirical measurements, not derivations from their own inputs. The 11.22x memory reduction and 2.23x speedup are reported against ZipVoice-Dialog and VibeVoice on the CoVoMix2 and OpenDialog test sets; quality is assessed with external WER, cpSIM, and UTMOS metrics. The masked flow-matching objective (Eq. 3) and auxiliary mel reconstruction loss (Eq. 5) are training losses, and the downsampling schedule [1,1,2,1,1] is selected via a controlled ablation on LibriSpeech-PC. Choosing hyperparameters (compression factor r=4, latent dimension D=100, loss weight lambda=0.5, schedule) by validation performance is standard model selection, not circular reasoning: none of the reported benchmark numbers are constructed from these choices by definition. The frozen autoencoder being trained on 3 s crops yet applied to 179 s dialogs is a legitimate extrapolation risk, but it is an empirical generalization concern, not a circularity. No self-citation is load-bearing, and no equation reduces a prediction to a fitted value. The derivation chain is self-contained against external baselines and data.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The system relies on hand-picked compression and schedule choices plus assumptions about representational sufficiency and metric validity. No formal theorem is proved and no new physical entities are introduced; the mask embedding z_mask is a standard learnable model component, not an invented external entity.

free parameters (5)
  • Temporal compression factor r = 4
    Chosen by hand; sets the 25 Hz latent rate. Central to the memory and speed gains; no sweep over other values is reported for the main dialog results.
  • Latent dimension D = 100
    Chosen by hand; balances decoder capacity against the sequence bottleneck in Section 3.1.
  • Auxiliary mel loss weight lambda = 0.5
    Set by hand in Section 4.2; balances L_vel and L_mel in Eq. 6. No sensitivity analysis is reported.
  • ZipFormer downsampling schedule = [1,1,2,1,1]
    Selected from ablations in Table 3, not derived from first principles; labeled 'Ours' in the paper.
  • Flow solver steps = 16
    Used for both ZipL-Dialog and ZipVoice-Dialog in Section 4.5; the quality/efficiency trade-off of step count is not swept.
axioms (5)
  • standard math Conditional flow matching with a linear interpolation path is a valid generative modeling objective.
    Invoked in Section 3.2, Eqs. 2-3, following Lipman et al. [7]; standard theory used as background.
  • domain assumption The 100 Hz, 100-bin Vocos-compatible mel-spectrogram is a sufficient acoustic representation for multi-turn dialog.
    All experiments train and evaluate on this representation (Sections 3.1, 4.2), so any information lost at the mel front-end cannot be recovered.
  • domain assumption A 4x time-compressed 25 Hz latent sequence retains enough phonetic and speaker detail for natural dialog.
    The entire approach depends on this representational capacity; the paper tests it empirically but never proves it, and Table 2 shows clear WER/cpSIM degradation.
  • domain assumption The autoencoder trained on 3-second random crops generalizes to 26-179 second dialog sequences when frozen.
    Section 4.2 trains the AE on 3s crops, then freezes it and uses it for long dialog; no analysis of long-sequence reconstruction drift is given.
  • domain assumption UTMOS is an adequate proxy for perceptual naturalness.
    The 'maintaining perceptual naturalness' claim in the abstract and conclusion is supported only by UTMOS; no human listening test or other perceptual measure is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 7713 in / 13052 out tokens · 131116 ms · 2026-08-02T06:27:54.245718+00:00 · methodology

0 comments
read the original abstract

Zero-shot dialog TTS benefits from flow-matching, but minute-scale generation on dense mel-spectrograms causes severe memory bottlenecks, often forcing unnatural chunked synthesis. We propose ZipL-Dialog, which shifts conditional flow-matching into a 4x time-compressed (25 Hz) latent space. To preserve acoustic fidelity under compression, we employ a deterministic mel autoencoder with auxiliary mel-domain supervision and optimize the ZipFormer's hierarchical downsampling schedule. Experiments show that ZipL-Dialog reduces maximum peak GPU memory by 11.22x and accelerates inference by 2.23x over the baseline, substantially lowering the memory footprint of single-pass multi-minute dialog synthesis while maintaining perceptual naturalness.

Figures

Figures reproduced from arXiv: 2607.12496 by Jihwan Kim, Nam Soo Kim.

Figure 1
Figure 1. Figure 1: Training overview of ZipL-Dialog. A ground-truth mel-spectrogram is encoded into a 4× time-compressed latent sequence (25 Hz). The ZipFormer decoder predicts the CFM velocity on the masked target region, conditioned on text embed￾dings and noisy latents. The model is jointly optimized with a latent velocity loss and an auxiliary mel-domain reconstruction loss. latent estimate Zˆ = (1 − m) ⊙ Z + m ⊙ (Zt + (… view at source ↗

Review history (2 revisions) →

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

31 extracted references · 10 linked inside Pith

  1. [1]

    Introduction Recent advances in neural speech synthesis have extended text- to-speech (TTS) beyond short, single-utterance synthesis toward long-formandmulti-turnconversational audio, including pod- casts, role-play dialogs, and interactive agents. In these settings, models must preserve naturalness, speaker and turn consistency, and contextual coherence ...

    Pith/arXiv arXiv 2026

  2. [2]

    Related Work 2.1. Flow Matching for Zero-Shot Dialog TTS Recent zero-shot TTS systems have increasingly adopted diffu- sion and flow-matching frameworks to enable non-autoregressive speech synthesis with strong naturalness and conditioning fi- delity. For conversational speech, models such as CoV oMix and CoV oMix2 [4, 5] have shown that conditional flow ...

  3. [3]

    Method ZipL-Dialog performs dialog generation in a temporally com- pressed latent acoustic space rather than directly in the frame- level mel domain. Given a frame-level mel-spectrogram, we first encode it into a lower-rate continuous latent sequence, ap- ply masked conditional flow matching in the latent domain, and finally decode the predicted latents b...

  4. [4]

    Experiments We conduct all experiments on English data, matching the lan- guage coverage of the large-scale training mixture and the evalu- ation protocols of prior dialog TTS benchmarks. 4.1. Training Datasets Backbone pretraining.Following the ZipV oice training recipe, we pretrain the backbone on a mixture of large-scale English speech corpora: HiFiTTS...

  5. [5]

    Conclusion We proposed ZipL-Dialog, an efficient latent conditional flow- matching framework for long-form spoken dialog synthesis. By moving generation to a 25 Hz deterministic latent space and re- designing the ZipFormer downsampling schedule, ZipL-Dialog substantially reduces the memory and runtime cost of minute- scale synthesis. On the CoV oMix2 and ...

  6. [6]

    All content and ideas remain the sole original work of the authors

    Generative AI Use Disclosure The authors utilized OpenAI ChatGPT and Google Gemini to improve the English phrasing, readability, and grammar of this manuscript. All content and ideas remain the sole original work of the authors

  7. [7]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,”arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  8. [8]

    Vibevoice technical report,

    Z. Peng, J. Yu, W. Wang, Y . Chang, Y . Sun, L. Dong, Y . Zhu, W. Xu, H. Bao, Z. Wanget al., “Vibevoice technical report,”arXiv preprint arXiv:2508.19205, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    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

    Pith/arXiv arXiv 2025

  10. [10]

    Soulx-podcast: Towards realistic 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 realistic long-form podcasts with dialectal and paralinguistic diversity,”arXiv preprint arXiv:2510.23541, 2025

    arXiv 2025

  11. [11]

    Covomix: Advancing zero-shot speech generation for human-like multi-talker conversations,

    L. Zhang, Y . Qian, L. Zhou, S. Liu, D. Wang, X. Wang, M. Yousefi, Y . Qian, J. Li, L. Heet al., “Covomix: Advancing zero-shot speech generation for human-like multi-talker conversations,”Advances in Neural Information Processing Systems, vol. 37, pp. 100 291– 100 317, 2024

    2024

  12. [12]

    Covomix2: Advancing zero-shot dialogue generation with fully non-autoregressive flow matching,

    L. Zhang, Y . Qian, X. Wang, M. Thakker, D. Wang, J. Yu, H. Wu, Y . Hu, J. Li, Y . Qianet al., “Covomix2: Advancing zero-shot dialogue generation with fully non-autoregressive flow matching,” arXiv preprint arXiv:2506.00885, 2025

    arXiv 2025

  13. [13]

    Zipvoice-dialog: Non- autoregressive spoken dialogue generation with flow matching,

    H. Zhu, W. Kang, L. Guo, Z. Yao, F. Kuang, W. Zhuang, Z. Li, Z. Han, D. Zhang, X. Zhanget al., “Zipvoice-dialog: Non- autoregressive spoken dialogue generation with flow matching,” arXiv preprint arXiv:2507.09318, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Libritts: A corpus derived from librispeech for text- to-speech,

    H. Zen, V . Dang, R. Clark, Y . Zhang, R. J. Weiss, Y . Jia, Z. Chen, and Y . Wu, “Libritts: A corpus derived from librispeech for text- to-speech,”arXiv preprint arXiv:1904.02882, 2019

    Pith/arXiv arXiv 1904

  15. [15]

    and uses acoustic prompts from LibriSpeech test-clean [16]. 4.4. Baseline Models We compare ZipL-Dialog against two strong baselines from complementary generative paradigms: • ZipVoice-Dialog (NAR baseline) [6]:the most closely related frame-level non-autoregressive dialog TTS baseline built on the same architectural family. • VibeVoice 1.5B (AR baseline)...

    arXiv 2034

  16. [16]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  17. [17]

    Flow matching in latent space,

    Q. Dao, H. Phung, B. Nguyen, and A. Tran, “Flow matching in latent space,”arXiv preprint arXiv:2307.08698, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    M3-tts: Multi-modal dit alignment & mel- latent for zero-shot high-fidelity speech synthesis,

    X. Wang, C. Qiang, R. Fu, Z. Wen, X. Liu, Y . Liu, Y . Liang, K. Yin, Y . Xie, H. Xieet al., “M3-tts: Multi-modal dit alignment & mel- latent for zero-shot high-fidelity speech synthesis,”arXiv preprint arXiv:2512.04720, 2025

    arXiv 2025

  19. [19]

    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,” arXiv preprint arXiv:2306.00814, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    Hifitts-2: A large-scale high bandwidth speech dataset,

    R. Langman, X. Yang, P. Neekhara, S. Hussain, E. Casanova, E. Bakhturina, and J. Li, “Hifitts-2: A large-scale high bandwidth speech dataset,”arXiv preprint arXiv:2506.04152, 2025

    arXiv 2025

  21. [21]

    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

  22. [22]

    Dailydialog: A manually labelled multi-turn dialogue dataset,

    Y . Li, H. Su, X. Shen, W. Li, Z. Cao, and S. Niu, “Dailydialog: A manually labelled multi-turn dialogue dataset,” inProceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2017, pp. 986–995

    2017

  23. [23]

    Librispeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: An ASR corpus based on public domain audio books,” in2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  24. [24]

    Parakeet: A natural sounding, conversational text-to-speech model,

    J. Darefsky, G. Zhu, and Z. Duan, “Parakeet: A natural sounding, conversational text-to-speech model,” 2024, blog post. [Online]. Available: https://jordandarefsky.com/blog/2024/parakeet/

    2024

  25. [25]

    Pyannote. audio: neural building blocks for speaker diarization,

    H. Bredin, R. Yin, J. M. Coria, G. Gelly, P. Korshunov, M. Lavechin, D. Fustes, H. Titeux, W. Bouaziz, and M.-P. Gill, “Pyannote. audio: neural building blocks for speaker diarization,” inICASSP 2020-2020 IEEE International conference on acous- tics, speech and signal processing (ICASSP). IEEE, 2020, pp. 7124–7128

    2020

  26. [26]

    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

  27. [27]

    Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,”arXiv preprint arXiv:2005.07143, 2020

    Pith/arXiv arXiv 2005

  28. [28]

    Utmos: Utokyo-sarulab system for voicemos chal- lenge 2022,

    T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari, “Utmos: Utokyo-sarulab system for voicemos chal- lenge 2022,”arXiv preprint arXiv:2204.02152, 2022

    Pith/arXiv arXiv 2022

  29. [29]

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

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. JianZhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025, pp. 6255–6271

    2025

  30. [30]

    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 understand- ing workshop (ASRU). IEEE, 2023, pp. 1–7

    2023

  31. [31]

    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 trans- actions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021