Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Enhancing Intelligibility for Generative Target Speech Extraction via Joint Optimization with Target Speaker ASR

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Jointly optimizing a generative target speech extraction model with a transcript-prediction loss yields extracted speech with both high perceptual quality and preserved semantic content.

desk verdict A credible joint-training ablation for Whisper-based generative TSE, but the central WER claim needs an independent ASR before I'd fully trust it. read the letter →

arxiv 2501.14477 v2 pith:4YACVMJ5 submitted 2025-01-24 eess.AS cs.SD

Hao Ma , Rujin Chen , Xiao-Lei Zhang , Ju Liu , Xuelong Li This is my paper · ORCID
classification eess.AScs.SD
keywords generativetargetspeechextractionWhisperflowmatchingintelligibilityspeakerASRjointoptimizationlow-rankadaptationseparation
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

The paper tries to show that a generative target speech extraction system can recover a target speaker's voice from a mixed recording without mangling what the speaker said. It builds the extractor on the pre-trained Whisper audio encoder, which is prompted to attend to the target speaker, and re-synthesizes the speech from continuous tokens by optimal-transport conditional flow matching. The new step is to train this whole pipeline with both the flow-matching objective and a transcript-prediction cross-entropy objective, so semantic content is supervised as well as acoustics. Experiments on Libri2Mix and WSJ0-2mix report better DNSMOS, lower word error rate, and higher SpeechBERTScore than discriminative and generative baselines, and the ablation shows that removing the transcript loss raises WER from 0.142 to 0.234.

What carries the argument

The machine is a shared target speech encoder built from the pre-trained Whisper audio encoder, adapted by LoRA and prompted by concatenating the target speaker embedding plus raw enrollment speech ahead of the mixture input (Eq. 2). Downstream, an optimal-transport conditional flow matching module (OT-CFM, Eq. 5) turns the extracted tokens into a mel-spectrogram which HiFiGAN turns into a waveform, while a text decoder (Eq. 6) turns the same tokens into a transcript under a cross-entropy loss (Eq. 7). The joint loss (Eq. 8) is what carries the claim: the text decoder is not an output branch used at inference, but a training-time supervisor that pushes the encoder to keep the semantic content that the transcript represents.

What would settle it

Run the same WhisperTSE models through an independently trained ASR system not derived from Whisper and collect human listening ratings on intelligibility. If the word error rate advantage over the w/o-joint-training model shrinks to noise, or listeners do not prefer the joint-trained outputs, the central claim that transcript supervision improves intelligibility would be refuted. A cheaper pre-registered check is to re-evaluate the published checkpoints on Libri2Mix-test-clean and compare the WER and SBS numbers against the Table II values of 0.142 and 0.907.

Watch

Extended reading notes

Core claim

The central claim is that intelligibility and perceptual quality in generative target speech extraction are not in conflict if the model is jointly optimized for both. The same target speech tokens extracted by the prompted Whisper encoder feed two branches: a flow-matching synthesizer that generates the mel-spectrogram, and a text decoder that predicts the transcript; the combined loss $L = L_{\mathrm{OT-CFM}} + L_{\mathrm{CE}}$ trains the encoder and synthesizer so that the generated audio is both natural and semantically accurate. On Libri2Mix-test-clean, WhisperTSE-S reaches WER 0.142 and SBS 0.907, while removing the cross-entropy term degrades WER to 0.234 and SBS to 0.891 (Table II), which the authors take as direct evidence that transcript supervision is what preserves meaning. The authors also report that the method outperforms the mask-based pBSRoformer and the token-LM baseline TSELM on quality and intelligibility metrics, while acknowledging that cosine similarity to the reference remains lower than for discriminative methods.

Load-bearing premise

The load-bearing premise is that the automatic metrics used—Whisper-small WER, DNSMOS, SpeechBERTScore, and CAM++ cosine similarity—faithfully capture what human listeners mean by intelligible and natural speech, since no human listening test is reported.

Editorial extensions

If this is right

  • A Whisper-based encoder with LoRA can serve as a target-speaker extractor, not just a recognizer, when prompted with enrollment speech.
  • Training-time transcript supervision can substitute for explicit semantic losses in generative speech separation, as shown by the WER and SBS drops when it is removed.
  • Generative TSE needs non-intrusive evaluation such as DNSMOS, WER, and SpeechBERTScore, because intrusive metrics like SNR are not meaningful after mel-based vocoder synthesis.
  • Cascading ASR with text-to-speech loses paralinguistic information, whereas jointly generating acoustics and semantics retains more of that information, as indicated by the cosine similarity results.
  • The proposed framework provides a concrete instance where a pre-trained ASR encoder is repurposed for speech generation without catastrophic forgetting of linguistic content.

Reading between the lines

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

  • A likely next test is whether the intelligibility gain holds when the evaluation ASR is not from the Whisper family, since the model itself uses a Whisper encoder and could be biased toward Whisper-compatible outputs.
  • The transcript-supervision idea could generalize to other generative speech tasks, such as speech enhancement or voice conversion, offering a generic meaning-preserving regularizer for models that currently optimize only acoustic losses.
  • Because LoRA rank 16 already captures most of the benefit, the Whisper-based TSE pipeline could be compressed further; the authors themselves note inference cost as a limitation, so distillation or pruning experiments would be the natural follow-up.
  • A human listening study that separately rates intelligibility and naturalness would sharpen the claim beyond the reported automatic metrics, and would test whether DNSMOS and SpeechBERTScore tell the same story as human perception.
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 / 6 minor

Summary. The paper proposes a generative target speech extraction (TSE) system built on a pre-trained Whisper encoder, adapted to a target speaker via a speaker embedding and raw enrollment speech, with LoRA fine-tuning. The extracted speech tokens are passed through two branches: an optimal-transport conditional flow matching (OT-CFM) module that synthesizes a mel-spectrogram (later inverted by HiFi-GAN), and a frozen Whisper text decoder that predicts the target transcript. The two branches are trained jointly with a sum of the flow-matching loss and a cross-entropy text loss (Eq. 8). The main claimed contribution is that this joint optimization improves the intelligibility of generative TSE output. Experiments on Libri2Mix and WSJ0-2mix report DNSMOS, WER, SpeechBERTScore, and cosine similarity, and an ablation in Table II shows that removing the text loss raises WER from 0.142 to 0.234 and lowers SpeechBERTScore from 0.907 to 0.891, supporting the intelligibility benefit of the joint loss.

Significance. If the reported intelligibility gain is genuine and transfers to independent ASR systems, the paper makes a useful contribution: it identifies a concrete failure mode of generative TSE (semantic distortion) and provides a simple, effective remedy by adding transcript supervision from a frozen Whisper decoder. The ablation in Table II is well designed and directly isolates the influence of the text loss, and the comparison against a discriminative baseline, a discrete-token generative baseline, and a cascaded pipeline is appropriate. The authors also provide audio samples on a demo page and evaluate three model sizes. The main weakness is that the primary intelligibility metric (Whisper-small WER) shares the Whisper model family with the semantic branch, and the abstract claims superiority over all baselines despite Table I showing that the cascading baseline has lower WER on Libri2Mix and the discriminative baseline has higher cosine similarity on both datasets. These issues, rather than the core derivation, are what currently prevent acceptance.

major comments (4)
  1. [Abstract; Table I; Section I] The abstract and Section I claim that the proposed method outperforms existing generative and discriminative baselines, but Table I does not support this as stated. On Libri2Mix, the Cascading baseline achieves WER 0.070 versus 0.142 for WhisperTSE-S, and pBSRoformer achieves cosine similarity 0.905 versus 0.749; on WSJ0-2mix, pBSRoformer again has higher cosine similarity (0.927 versus 0.814). Since no metric priority or trade-off is defined, please replace the global claim with a qualified one that names the metrics on which the method is superior (DNSMOS, WER, SBS) and the metric on which it is not (cosine similarity).
  2. [Section III-B; Section II-C; Eq. (6)-(7); Table II] The key intelligibility evidence uses Whisper-small to compute WER, while the model's semantic branch is a frozen Whisper text decoder and the cross-entropy loss L_CE in Eq. (7) explicitly trains the encoder tokens H_i to be decodable by that decoder. The with-L_CE model is therefore optimized toward Whisper-specific decoding regularities, so the Table II gap (WER 0.142 vs 0.234) may overstate a general intelligibility improvement. Please re-evaluate the ablation with an ASR system from outside the Whisper family (e.g., a CTC/HuBERT-based or Kaldi recognizer) and report those numbers; a small human listening test for intelligibility would further settle the claim.
  3. [Table II; Section III-D3] The interpretation that joint training makes the model 'overly focused on acoustic details, leading to a loss in its semantic modeling capability' (Section III-D3) is not supported by Table II: removing L_CE increases cosine similarity from 0.749 to 0.793 while worsening WER and SBS. The higher cosine similarity without joint training indicates better preservation of non-linguistic/speaker information, not greater focus on acoustic details in a semantic sense. The authors should re-frame this trade-off and consider reporting it as a known limitation rather than as a side effect of semantic modeling.
  4. [Section III-B; Section IV] The paper's central 'high intelligibility and perceptual quality' claim rests entirely on non-intrusive metrics (DNSMOS, WER by Whisper-small, SpeechBERTScore, cosine similarity). Section III-B justifies excluding intrusive metrics such as SNR, but no human listening test or second independent perceptual metric is reported. Given that DNSMOS can miss certain vocoder artifacts and that the WER metric is family-biased as noted above, the perceptual-quality claim is not yet established at the level claimed. I ask for at least a small listening test or a clear statement that the claims are limited to these proxy metrics.
minor comments (6)
  1. [Section II-D; Eq. (2)] The trainable-parameter description in Section II-D does not mention the affine layer W in Eq. (2); please state explicitly whether W and the additional positional encoding Pos' are trained or frozen.
  2. [Fig. 2] The axes lack labels for the displayed normalized metrics; please add a legend and describe the normalization (e.g., min-max per metric) in the caption.
  3. [Section III-C] The LoRA rank is selected on WhisperTSE-S and then fixed at 16 for -M and -L; please justify that the optimal rank transfers across model sizes, or provide rank sweeps for at least one larger model.
  4. [Tables I and II] Only point estimates are reported; adding confidence intervals or significance tests (e.g., bootstrap over utterances) would make comparisons such as WhisperTSE-M versus WhisperTSE-L (WER 0.109 vs 0.081) more interpretable.
  5. [Section II-B; Eq. (5)] The hyperparameter sigma in Eq. (5) is described only as 'small'; please report its value and the number of flow sampling steps used at inference.
  6. [Abstract] The demo URL in the visible text contains a space ('GenerativeTSE demo/'); please correct it to match the working link 'GenerativeTSE_demo'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the joint-training gain is an empirical comparison, and the only self-citation is background, not load-bearing.

full rationale

The paper's derivation is a training-objective construction, not a fitted-prediction loop. Eq. (8) defines L = L_OT-CFM + L_CE, where L_CE (Eq. 7) supervises the Whisper text decoder on transcript tokens and L_OT-CFM (Eq. 5) supervises the flow-based mel synthesizer; neither term is defined in terms of the evaluation quantities WER, DNSMOS, SBS, or cosine similarity. The load-bearing ablation (Table II) compares models trained with and without L_CE and measures WER with an external pre-trained Whisper-small checkpoint on Libri2Mix-test-clean; there is no parameter fitted to that test metric, and the reported WER gap is an empirically contingent outcome rather than a construction identity. The only self-citation is [12] in the introductory list of target-speaker ASR works; the actual prompting design is attributed to [13] (Meng et al.), a non-overlapping group, and the capability claim is independently supported by [13], [14]. The fact that the model uses Whisper-family components and is scored by a Whisper-family ASR is a legitimate measurement and validity limitation, since the intelligibility metric is not fully independent of the model's inductive bias, but it does not make the claim true by definition, and no equation in the paper reduces to its own output. Therefore no circular step meets the evidentiary bar.

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

The model introduces no theoretical free parameters beyond standard training hyperparameters and borrowed pretrained components. The only user-selected hyperparameter with a sensitivity study is LoRA rank. The evaluation rests on non-intrusive and ASR-based metrics because the authors deem intrusive metrics inapplicable; this is the largest unvalidated premise. No invented physical or architectural entities are introduced beyond the assembled Whisper-plus-flow pipeline.

free parameters (1)
  • LoRA rank r = 16
    Chosen by sweeping {0, 4, 8, 16, 32, 64, full} on Libri2Mix-dev-clean (Fig. 2); all final models use r=16. This is a manually selected hyperparameter, not derived.
assumptions (4)
  • standard math Optimal transport conditional flow matching (OT-CFM) constructs a valid probability path from a Gaussian prior to the target speech distribution.
    Eq. (5) relies on the flow-matching framework of [17]; the paper does not re-derive it.
  • domain assumption A Whisper audio encoder pre-trained on single-talker ASR can be adapted with LoRA and prompt concatenation to emit tokens that isolate the target speaker from a two-talker mixture.
    Eq. (2), Section II-A; this is the central transfer assumption and is tested only indirectly through downstream metrics.
  • domain assumption DNSMOS, Whisper-small WER, SpeechBERTScore, and CAM++ cosine similarity are faithful proxies for perceived quality, intelligibility, and timbre preservation.
    Section III-B defines the evaluation; no human listening test or standard intrusive metrics are reported.
  • domain assumption Intrusive metrics such as SNR are not applicable to vocoder-based generative outputs, so their absence is justified.
    Section III-B, citing [6], [7], [32]; if this is wrong, the comparison omits a standard objective quality check.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Intelligibility for Generative Target Speech Extraction via Joint Optimization with Target Speaker ASR." pith.science (2026). https://pith.science/paper/4YACVMJ5

@misc{pith2026250114477,
  author       = {Pith},
  title        = {Pith review of: Enhancing Intelligibility for Generative Target Speech Extraction via Joint Optimization with Target Speaker ASR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4YACVMJ5}},
  note         = {Machine review of arXiv:2501.14477}
}
read the original abstract

Target speech extraction (TSE) isolates the speech of a specific speaker from a multi-talker overlapped speech mixture. Most existing TSE models rely on discriminative methods, typically predicting a time-frequency spectrogram mask for the target speech. However, imperfections in these masks often result in over-/under-suppression of target/non-target speech, degrading perceptual quality. Generative methods, by contrast, re-synthesize target speech based on the mixture and target speaker cues, achieving superior perceptual quality. Nevertheless, these methods often overlook speech intelligibility, leading to alterations or loss of semantic content in the re-synthesized speech. Inspired by the Whisper model's success in target speaker ASR, we propose a generative TSE framework based on the pre-trained Whisper model to address the above issues. This framework integrates semantic modeling with flow-based acoustic modeling to achieve both high intelligibility and perceptual quality. Results from multiple benchmarks demonstrate that the proposed method outperforms existing generative and discriminative baselines. We present speech samples on https://aisaka0v0.github.io/GenerativeTSE_demo/.

Figures

Figures reproduced from arXiv: 2501.14477 by the authors.

Figure 1
Figure 1. Overview of the proposed method. The solid line represents the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Normalized metrics versus LoRA rank. All metrics are evaluated with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Spectrogram Visualization superior perceptual quality. Moreover, the speech extracted using our model, based on its energy distribution in the time-frequency spectrogram, shows no perceptible distortion compared to the ground truth. This suggests that both the linguistic and non-linguistic information in the original speech mixture is effectively preserved. IV. DISCUSSION In this work, we propose a generative target… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. FlowTSE: Target Speaker Extraction with Flow Matching

    eess.AS 2025-05 conditional novelty 6.0 of 10

    Conditional flow matching on mel-spectrograms with a phase-conditioned vocoder matches or beats published TSE baselines on Libri2Mix.

  2. SoloSpeech: Enhancing Intelligibility and Quality in Target Speech Extraction through a Cascaded Generative Pipeline

    eess.AS 2025-05 conditional novelty 5.0 of 10

    A cascaded pipeline of audio compression, latent diffusion extraction, and generative correction achieves state-of-the-art target speech extraction quality and intelligibility on Libri2Mix and out-of-domain data.

Reference graph

Works this paper leans on

34 extracted references · 23 canonical work pages · cited by 2 Pith papers

  1. [1]

    Supervised speech separation based on deep learning: An overview,

    D. Wang and J. Chen, “Supervised speech separation based on deep learning: An overview,” IEEE/ACM Trans. Audio, Speech, Lang. Pro- cess., vol. 26, no. 10, pp. 1702–1726, 2018

    2018

  2. [2]

    SpeakerBeam: Speaker aware neural network for target speaker extraction in speech mixtures,

    K. ˇZmol´ıkov´a, M. Delcroix, K. Kinoshita, T. Ochiai, T. Nakatani, L. Burget, and J. ˇCernock´y, “SpeakerBeam: Speaker aware neural network for target speaker extraction in speech mixtures,” IEEE J. Sel. Topics Signal Process. , vol. 13, no. 4, pp. 800–814, 2019

    work page 2019

  3. [3]

    Improving speaker discrimination of target speech extraction with time-domain speakerbeam,

    M. Delcroix, T. Ochiai, K. Zmolikova, K. Kinoshita, N. Tawara, T. Nakatani, and S. Araki, “Improving speaker discrimination of target speech extraction with time-domain speakerbeam,” in Proc. ICASSP , 2020, pp. 691–695

    work page 2020

  4. [4]

    SpEx: Multi-scale time domain speaker extraction network,

    C. Xu, W. Rao, E. S. Chng, and H. Li, “SpEx: Multi-scale time domain speaker extraction network,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 28, pp. 1370–1384, 2020

    2020

  5. [5]

    Target speech extraction with conditional diffusion model,

    N. Kamo, M. Delcroix, and T. Nakatani, “Target speech extraction with conditional diffusion model,” in Proc. INTERSPEECH, 2023, pp. 176– 180

    work page 2023

  6. [6]

    Generation- based target speech extraction with speech discretization and vocoder,

    L. Yu, W. Zhang, C. Du, L. Zhang, Z. Liang, and Y . Qian, “Generation- based target speech extraction with speech discretization and vocoder,” in Proc. ICASSP, 2024, pp. 12 612–12 616

    work page 2024

  7. [7]

    TSELM: Target speaker extraction using discrete tokens and language models,

    B. Tang, B. Zeng, and M. Li, “TSELM: Target speaker extraction using discrete tokens and language models,” arXiv preprint arXiv:2409.07841, 2024

    arXiv 2024

  8. [8]

    X-Sepformer: End-to-end speaker extraction network with explicit optimization on speaker confusion,

    K. Liu, Z. Du, X. Wan, and H. Zhou, “X-Sepformer: End-to-end speaker extraction network with explicit optimization on speaker confusion,” in Proc. ICASSP, 2023, pp. 1–5

    work page 2023

Show all 34 references
  1. [9]

    Personalized speech enhancement combining band-split rnn and speaker attentive module,

    X. Le, L. Chen, C. He, Y . Guo, C. Chen, X. Xia, and J. Lu, “Personalized speech enhancement combining band-split rnn and speaker attentive module,” in Proc. ICASSP, 2023, pp. 1–2

    2023

  2. [10]

    Multi- level speaker representation for target speaker extraction,

    K. Zhang, J. Li, S. Wang, Y . Wei, Y . Wang, Y . Wang, and H. Li, “Multi- level speaker representation for target speaker extraction,” arXiv preprint arXiv:2410.16059, 2024

    2024 arXiv

  3. [11]

    Hierarchical speaker representation for target speaker extraction,

    S. He, H. Zhang, W. Rao, K. Zhang, Y . Ju, Y . Yang, and X. Zhang, “Hierarchical speaker representation for target speaker extraction,” in Proc. ICASSP, 2024, pp. 10 361–10 365

    2024

  4. [12]

    Extending Whisper with prompt tuning to target-speaker ASR,

    H. Ma, Z. Peng, M. Shao, J. Li, and J. Liu, “Extending Whisper with prompt tuning to target-speaker ASR,” in Proc. ICASSP, 2024

    2024

  5. [13]

    Empowering Whisper as a joint multi-talker and target-talker speech recognition system,

    L. Meng, J. Kang, Y . Wang, Z. Jin, X. Wu, X. Liu, and H. Meng, “Empowering Whisper as a joint multi-talker and target-talker speech recognition system,” in Proc. INTERSPEECH, 2024

    2024

  6. [14]

    SQ-Whisper: Speaker-querying based Whisper model for target-speaker ASR,

    P. Guo, X. Chang, H. Lv, S. Watanabe, and L. Xie, “SQ-Whisper: Speaker-querying based Whisper model for target-speaker ASR,” IEEE/ACM Trans. Audio, Speech, Lang. Process. , pp. 1–11, 2024

    2024

  7. [15]

    Robust speech recognition via large-scale weak super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” in Proc. ICML, 2023, pp. 28 492–28 518

    2023

  8. [16]

    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,” ICLR, 2022

    2022

  9. [17]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” ICLR, 2023

    2023

  10. [18]

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

    S. Mehta, R. Tu, J. Beskow, ´E. Sz ´ekely, and G. E. Henter, “Matcha- TTS: A fast TTS architecture with conditional flow matching,” in Proc. ICASSP, 2024

    2024

  11. [19]

    CosyV oice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens,

    Z. Du, Q. Chen, S. Zhang, K. Hu, H. Lu, Y . Yang, H. Hu, S. Zheng, Y . Gu, Z. Ma et al. , “CosyV oice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens,” arXiv preprint arXiv:2407.05407, 2024

    2024 arXiv

  12. [20]

    LibriMix: An open-source dataset for generalizable speech separation,

    J. Cosentino, M. Pariente, S. Cornell, A. Deleforge, and E. Vincent, “LibriMix: An open-source dataset for generalizable speech separation,” 2020

    2020

  13. [21]

    Deep clustering: Discriminative embeddings for segmentation and separation,

    J. R. Hershey, Z. Chen, J. Le Roux, and S. Watanabe, “Deep clustering: Discriminative embeddings for segmentation and separation,” in Proc. ICASSP, 2016, pp. 31–35

    2016

  14. [22]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proc. NeurIPS, vol. 30, 2017

    2017

  15. [23]

    Drop the beat! freestyler for accompaniment conditioned rapping voice generation,

    Z. Ning, S. Wang, Y . Jiang, J. Yao, L. He, S. Pan, J. Ding, and L. Xie, “Drop the beat! freestyler for accompaniment conditioned rapping voice generation,” arXiv preprint arXiv:2408.15474 , 2024

    2024 arXiv

  16. [24]

    StableVC: Style controllable zero-shot voice conversion with conditional flow matching,

    J. Yao, Y . Yan, Y . Pan, Z. Ning, J. Ye, H. Zhou, and L. Xie, “StableVC: Style controllable zero-shot voice conversion with conditional flow matching,” arXiv preprint arXiv:2412.04724 , 2024

    2024 arXiv

  17. [25]

    HiFi-GAN: Generative adversarial net- works for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: Generative adversarial net- works for efficient and high fidelity speech synthesis,” in Proc. NeurIPS, vol. 33, 2020, pp. 17 022–17 033

    2020

  18. [26]

    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,” in Proc. ICASSP , 2015, pp. 5206–5210

    2015

  19. [27]

    Adapting self- supervised models to multi-talker speech recognition using speaker embeddings,

    Z. Huang, D. Raj, P. Garc ´ıa, and S. Khudanpur, “Adapting self- supervised models to multi-talker speech recognition using speaker embeddings,” in Proc. ICASSP, 2023

    2023

  20. [28]

    Single channel speech separation with constrained utterance level permutation invariant training using grid lstm,

    C. Xu, W. Rao, X. Xiao, E. S. Chng, and H. Li, “Single channel speech separation with constrained utterance level permutation invariant training using grid lstm,” in Proc. ICASSP, 2018, pp. 6–10

    2018

  21. [29]

    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,” in Proc. ICASSP, 2022, pp. 886–890

    2022

  22. [30]

    Speechbertscore: Reference-aware automatic evaluation of speech generation leveraging nlp evaluation metrics,

    T. Saeki, S. Maiti, S. Takamichi, S. Watanabe, and H. Saruwatari, “Speechbertscore: Reference-aware automatic evaluation of speech generation leveraging nlp evaluation metrics,” arXiv preprint arXiv:2401.16812, 2024

    2024 arXiv

  23. [31]

    CAM++: A fast and efficient network for speaker verification using context-aware masking,

    H. Wang, S. Zheng, Y . Chen, L. Cheng, and Q. Chen, “CAM++: A fast and efficient network for speaker verification using context-aware masking,” in Proc. INTERSPEECH, 2023, pp. 5301–5305

    2023

  24. [32]

    SELM: Speech enhancement using discrete tokens and language mod- els,

    Z. Wang, X. Zhu, Z. Zhang, Y . Lv, N. Jiang, G. Zhao, and L. Xie, “SELM: Speech enhancement using discrete tokens and language mod- els,” in Proc. ICASSP, 2024, pp. 11 561–11 565

    2024

  25. [33]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” ICLR, 2019

    2019

  26. [34]

    Music source separation with band-split rope transformer,

    W.-T. Lu, J.-C. Wang, Q. Kong, and Y .-N. Hung, “Music source separation with band-split rope transformer,” in Proc. ICASSP , 2024, pp. 481–485

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.