Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

FlowTSE: Target Speaker Extraction with Flow Matching

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A flow-matching model over mel-spectrograms matches or outperforms strong target-speaker-extraction baselines with a simpler pipeline.

desk verdict FlowTSE is a clean flow-matching TSE system with competitive Libri2Mix numbers, but the headline 'matches or outperforms' needs the missing strong baselines before it can be taken at face value. read the letter →

arxiv 2505.14465 v1 pith:LUMRRNR4 submitted 2025-05-20 eess.AS cs.LGcs.SD

Aviv Navon , Aviv Shamsian , Yael Segal-Feldman , Neta Glazer , Gil Hetz , Joseph Keshet This is my paper · ORCID
classification eess.AScs.LGcs.SD
keywords targetspeakerextractionflowmatchingconditionalspeechenhancementmel-spectrogramphase-conditionedvocodergenerativeseparationLibri2Mix
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

FlowTSE asks whether target speaker extraction can be done by a generative model over mel-spectrograms, conditioned only on an enrollment clip and the mixed signal, rather than by complex pipelines with pretrained discriminative or ASR components. The paper's central claim is that conditional flow matching makes this work: on Libri2Mix clean and noisy mixtures, FlowTSE matches or outperforms both discriminative and generative baselines, including the highest PESQ among the compared methods, and it also improves speech enhancement without any retraining. The authors further claim that a phase-conditioned vocoder, which blends the mixed signal's complex STFT with the predicted one, can repair the waveform-level alignment that mel-spectrograms discard. If true, this would make generative TSE simpler to train and deploy while preserving perceptual quality.

What carries the argument

The load-bearing object is the conditional flow matching (CFM) objective: the model is trained to regress the velocity field $(x_1 - x_0)$ along linear interpolation paths between Gaussian noise and the clean target mel, and at inference the learned field is integrated from $t=0$ to $t=1$. The conditioning is the pair of mel-spectrograms, enrollment and mixture, concatenated along time with asymmetric attention, so mixed tokens attend to enrollment tokens but not vice versa. The phase-conditioned vocoder is the second mechanism: it modifies the Vocos architecture with cross-attention on the mixed signal's complex STFT and a head that forms a complex linear combination $\alpha \odot s_m + \beta \odot s_p$ between the mixed and predicted STFT, with per-bin learned weights, motivated by alignment of the generated audio to the mixture. This component is trainable independently with an SI-SDR objective and converts the mel output into a waveform with phase informed by the mixture.

What would settle it

Re-run the Table 1 evaluation on the actual vocoded waveforms and compute SI-SDR against the clean reference; with the standard vocoder the paper reports -25.64 dB, so if that near-zero alignment persists while PESQ improves, the perceptual-quality claim would not reflect waveform fidelity.

Watch

Extended reading notes

Core claim

The discovery is that a flow-matching model, trained with the conditional flow matching objective and a diffusion-transformer backbone, can extract a target speaker directly in the mel-spectrogram domain, using concatenated enrollment and mixture mel frames as conditioning with one-way attention from mixture to enrollment. The model regresses a velocity field from noise to the clean target mel, then integrates it at inference with an ODE solver, and converts the resulting mel to a waveform with a vocoder. On Libri2Mix, this simple pipeline reports the best PESQ among compared methods on both noisy and clean splits, the best OVRL and DNSMOS on clean, and near-best speaker similarity, with word error rates comparable to a much larger ASR-optimized baseline. The second claimed result is that the same pretrained model, without additional training, improves single-speaker speech enhancement. A third result is that a phase-conditioned vocoder raises SI-SDR from -25.64 dB to 10.81 dB, at the cost of lower PESQ and OVRL, showing a tradeoff between waveform alignment and perceptual quality.

Load-bearing premise

The strongest claims assume that a vocoder-inverted mel-spectrogram faithfully represents the extracted speech, yet the paper's own Table 3 shows that the standard vocoder yields essentially zero waveform alignment (SI-SDR of -25.64 dB) and the phase-correcting vocoder lowers PESQ and OVRL.

Editorial extensions

If this is right

  • FlowTSE's main results are produced by a pipeline that needs no pretrained discriminative TSE model, no SSL embedding, and no ASR supervision, so the same recipe should be reproducible on other datasets and languages.
  • Because the enrollment condition is a mel-spectrogram, the same model can be pointed at a speaker by swapping the enrollment clip without retraining.
  • The single-speaker enhancement result means that a model trained for extraction can double as an enhancer by feeding the mixture as its own enrollment, removing the need for a separate enhancement system.
  • The phase-conditioned vocoder is a plug-in that trades a small drop in PESQ and OVRL for a large gain in SI-SDR, making the approach usable where downstream waveform-level processing matters.
  • The reported low WER on Libri2Mix clean indicates that the extracted mel output, once vocoded, is intelligible enough for ASR, comparable to a much larger model that was explicitly trained with ASR supervision.

Reading between the lines

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

  • An implication the authors leave implicit is that, since all main metrics are computed on vocoder-inverted waveforms, a substantial part of the perceptual gain may be attributable to the vocoder rather than to the extraction module; readers who care about waveform-level fidelity should weigh Table 3 before adopting the method.
  • The asymmetric attention design suggests a natural extension to multi-speaker attribution: instead of one enrollment clip, the model could attend to a bank of enrollment tokens for different speakers and extract each in a single pass.
  • The phase-conditioned vocoder idea, mixing the input mixture's complex STFT with a predicted STFT via per-bin weights, is general and could be applied to other mel-based generative speech models that must align to a reference signal.
  • A testable extension is to evaluate FlowTSE on speakers unseen during training, since the enrollment clip is the only speaker identity signal and the model never sees speaker labels.
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

3 major / 5 minor

Summary. The paper proposes FlowTSE, a target speaker extraction method based on conditional flow matching. The model takes enrollment and mixture audio as mel-spectrograms, concatenates them along time, and generates the target speaker's clean mel-spectrogram using a flow-matching objective; a separate contribution is a phase-conditioned Vocos vocoder that leverages the complex STFT of the mixture to improve phase reconstruction. Experiments on Libri2Mix (min split, clean and noisy) report PESQ, ESTOI, OVRL, DNSMOS, and speaker similarity, with FlowTSE obtaining the best PESQ among the compared methods and competitive scores on other metrics. The paper also shows that the same pretrained model can be applied to single-speaker speech enhancement without retraining, and it evaluates the phase-conditioned vocoder in Table 3.

Significance. If the central claim holds, FlowTSE is a useful contribution: it is a comparatively simple generative TSE pipeline that avoids large pretrained ASR or SSL models, and it demonstrates that mel-domain flow matching can produce competitive perceptual quality. The paper is also transparent about the phase-vocoder trade-off in Table 3, and the zero-shot speech enhancement result is a nice bonus. However, the significance is currently limited by incomplete baseline coverage: several strong baselines listed in Section 4.1 do not appear in the main comparison table, and the waveform-level quality gap shown in Table 3 is discussed only partially.

major comments (3)
  1. [Section 4.1, Table 1] The baselines pBSRoformer, Whisper-TSE, and TSELM are introduced in Section 4.1 as comparison systems, but none of them appears in Table 1; SKIM appears only for PESQ and OVRL on mix-both. The only evidence involving these systems is Figure 2, which reports WER on the max split of mix-clean for WhisperTSE-S/L and pBSRoformer, not the PESQ/ESTOI/OVRL/DNSMOS/SIM suite used elsewhere. Since the abstract's central claim is that FlowTSE "matches or outperforms strong baselines," this claim is currently supported only against a selected subset of the listed baselines. Please add the missing baselines on the same metrics and splits, or explicitly restrict the claim to the evaluated systems.
  2. [Section 4.2.3, Table 3] The main results in Tables 1 and 2 are computed on waveforms obtained by inverting the generated mel-spectrogram with the standard Vocos vocoder, which Table 3 reports as SI-SDR = -25.64 dB, meaning essentially no waveform-level alignment. The proposed phase-conditioned vocoder recovers SI-SDR (10.81 dB) but lowers PESQ from 2.58 to 2.48 and OVRL from 3.27 to 3.14. The paper should state explicitly which vocoder configuration underlies each reported metric and should discuss the implications of this trade-off, because a reader who needs waveform-level fidelity would choose the phase vocoder and incur the perceptual degradation.
  3. [Section 4.1, Tables 1 and 2] Several baseline numbers are taken from earlier papers (NCSN++, DPCCN, and DiffSep from [2]; pBSRoformer from [12]) rather than re-evaluated under a common protocol, and no error bars or significance tests are reported. The performance differences supporting the "outperforms" wording are sometimes small (e.g., PESQ 1.86 vs. 1.74 for DPCCN in mix-both, and ESTOI ties in Table 2), so without variance estimates or paired significance tests the comparative claim is not robust. Please provide confidence intervals or significance tests for the main comparisons, or soften the wording to "comparable to."
minor comments (5)
  1. [Section 1] In the last paragraph, "conditioning the vocoder on the comple STFT" contains a typo: "comple" should be "complex."
  2. [Section 3.3] "While a simple recontract phase is sufficient for most applications" should read "reconstruct phase," and in Section 3.2 "train velocity field" should be "trained velocity field."
  3. [Table 2] The header "Mixure" is a typo and should be "Mixture."
  4. [Section 4.1] The min and max configurations of Libri2Mix are mentioned but not defined; please add a sentence or cite the dataset description for these settings.
  5. [Figure 2] The figure would be more useful if the exact WER numbers were also given in the text or caption, since the differences between FlowTSE and WhisperTSE-L appear small.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FlowTSE's claims are empirical and independently benchmarked; no load-bearing self-citation or definitional reduction found.

full rationale

FlowTSE is an empirical systems paper whose central claim is that a flow-matching model conditioned on enrollment and mixture mel-spectrograms matches or outperforms strong baselines. The derivation chain is not circular: the training objective is the standard conditional flow matching loss with the target clean mel as x1 and Gaussian noise as x0 (Eq. 1), and the model output is not fed back into its own conditioning. The enrollment and mixture inputs are external audio representations, not derived from the predicted output. Baselines are external methods, with results taken from other papers ([2], [12]), and the evaluation uses standard public benchmarks (Libri2Mix, WHAM!). The phase-conditioned vocoder conditions on the mixed signal's complex STFT and the clean mel, not on the model's own output, and it is trained with an SI-SDR objective; while Table 3 shows that this vocoder trades off perceptual metrics for SI-SDR, that is an empirical limitation rather than circular reasoning. The paper does not rely on self-citations, uniqueness theorems, or definitions that presuppose the target result. Concerns about omitted baselines or the poor SI-SDR of standard Vocos inversion are correctness risks, not circularity, and do not affect the circularity score. Therefore, the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper is an empirical deep learning system. It introduces no invented entities and fits no physical constants; its free parameters are standard training and architecture choices adopted from prior flow matching and vocoder work. The main unstated premises are that mel-spectrogram inversion is faithful enough for the reported quality metrics, that Libri2Mix/WHAM! represent the TSE task, and that Whisper-small WER is an acceptable intelligibility proxy.

free parameters (5)
  • Source distribution variance = 1 (N(0,I))
    The flow matching source p0 is a standard unit Gaussian in mel-feature space; its scale is not tuned for TSE.
  • Training SNR range = U[-5, 5] dB
    Mixture and noise SNRs are sampled from this range during training; chosen by the authors without optimization.
  • Noise insertion probability = 0.75
    WHAM! noise is added to training mixtures with probability 0.75; a hand-set augmentation choice.
  • Model architecture hyperparameters = 22 layers, 16 heads, dim 1024, FFN 2048
    Copied from F5-TTS [19]; not ablated or tuned for TSE.
  • Phase vocoder combination coefficients alpha and beta = learned per time-frequency bin
    Trained with the SI-SDR objective in the modified iSTFT head; they are model parameters, not externally fitted constants.
assumptions (5)
  • standard math Conditional flow matching training objective equals the marginal flow matching objective (Lipman et al., 2022).
    Invoked in Section 3.1 to justify optimizing L_CFM over paired x0 and x1; this is a published theorem the paper relies on without proof.
  • domain assumption Mel-spectrogram representation plus vocoder inversion is sufficient for high-quality target speaker extraction.
    The entire pipeline operates in the mel domain and discards phase by construction (Section 3.3); the reported quality metrics are computed after vocoder inversion.
  • domain assumption Libri2Mix and WHAM! are representative benchmarks for TSE and speech enhancement.
    Evaluation in Section 4 is limited to Libri2Mix min/max splits with WHAM! noise; generalization to other domains is not demonstrated.
  • domain assumption Whisper-small word error rate is a valid intelligibility proxy.
    Figure 2 reports WER from a pretrained ASR model without confidence intervals; no human listening test is included.
  • domain assumption Pretrained Vocos weights provide a reasonable default mel-to-waveform inverter.
    Main results use standard Vocos inversion; Table 3 quantifies that this inverter achieves very poor SI-SDR (-25.64 dB), so the assumption is only safe for perceptual metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlowTSE: Target Speaker Extraction with Flow Matching." pith.science (2026). https://pith.science/paper/LUMRRNR4

@misc{pith2026250514465,
  author       = {Pith},
  title        = {Pith review of: FlowTSE: Target Speaker Extraction with Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LUMRRNR4}},
  note         = {Machine review of arXiv:2505.14465}
}
read the original abstract

Target speaker extraction (TSE) aims to isolate a specific speaker's speech from a mixture using speaker enrollment as a reference. While most existing approaches are discriminative, recent generative methods for TSE achieve strong results. However, generative methods for TSE remain underexplored, with most existing approaches relying on complex pipelines and pretrained components, leading to computational overhead. In this work, we present FlowTSE, a simple yet effective TSE approach based on conditional flow matching. Our model receives an enrollment audio sample and a mixed speech signal, both represented as mel-spectrograms, with the objective of extracting the target speaker's clean speech. Furthermore, for tasks where phase reconstruction is crucial, we propose a novel vocoder conditioned on the complex STFT of the mixed signal, enabling improved phase estimation. Experimental results on standard TSE benchmarks show that FlowTSE matches or outperforms strong baselines.

Figures

Figures reproduced from arXiv: 2505.14465 by the authors.

Figure 1
Figure 1. FlowTSE architecture. The model is conditioned on the mixed and enrollment mel. However, current generative models also have their limitations. Many existing methods rely on complex pipelines involving dis￾criminative TSE models [2], or incorporating pretrained mod￾els such as Self-Supervised Learning (SSL) models [6, 11] or ASRs [12], which complicates their training and deploy￾ment. Additionally, several approache… view at source ↗
Figure 2
Figure 2. Comparison of word error rate evaluated on the Libri2Mix clean with max split. FlowTSE outperforms the WhisperTSE-S and pBSRoformer baselines while achieving a WER comparable to WhisperTSE-L, despite the latter being a significantly larger model. sists of 22 layers, with 16 attention heads, and an embedding di￾mension of 1024, followed by a 2048-dimensional feed-forward network. We train our model for 100 epochs wit… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. UmbraTTS: Adapting Text-to-Speech to Environmental Contexts with Flow Matching

    cs.SD 2025-06 conditional novelty 5.0 of 10

    UmbraTTS jointly synthesizes speech and environmental audio via conditional flow matching, conditioned on text and acoustic context, with controllable background volume.

Reference graph

Works this paper leans on

40 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [2]

    SE methods are generally categorized as discriminative or generative

    Related Work TSE falls under the broader category of Speech Enhancement (SE), which aims to recover clean speech from an input au- dio corrupted by noise. SE methods are generally categorized as discriminative or generative. Discriminative approaches di- rectly map noisy inputs to clean outputs, operating either in the frequency domain [4] or in the time ...

    work page

  2. [12]

    Speech enhancement and dereverberation with diffusion-based generative models,

    J. Richter, S. Welker, J.-M. Lemercier, B. Lay, and T. Gerkmann, “Speech enhancement and dereverberation with diffusion-based generative models,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2351–2364, 2023

    2023

  3. [1]

    TSE has gained significant attention for its role in improving speech clarity in communication systems and enhancing hearing aids for indi- viduals with hearing impairments [1, 2]

    Introduction Target Speaker Extraction (TSE) focuses on extracting the clean speech of a specific speaker from an audio mixture containing interfering speakers and background noise. TSE has gained significant attention for its role in improving speech clarity in communication systems and enhancing hearing aids for indi- viduals with hearing impairments [1...

    arXiv 2025

  4. [3]

    Our method maintains a simple pipeline that does not rely on external pretrained models besides the vocoder, which converts mel spectrograms into audio

    Method In this section, we present FlowTSE, our flow matching-based approach for TSE. Our method maintains a simple pipeline that does not rely on external pretrained models besides the vocoder, which converts mel spectrograms into audio. Additionally, we propose a simple yet effective modification to vocoder models, specifically V ocos [18], to enhance p...

    work page

  5. [4]

    Experimental Setup Datasets

    Experiments 4.1. Experimental Setup Datasets. We follow [12] to train our model using the Lib- riSpeech dataset [24]. Specifically, we use the train clean 100 and train clean 360 subsets. During training, we sample two random samples together with a noise sample from the WHAM! [25] dataset. The speech utterances are mixed with an SNR ratio sampled from U ...

    work page

  6. [5]

    We leverage recent advancements in flow matching to condition our model on the target speaker’s audio

    Discussion In this work, we introduce FlowTSE, a novel generative ap- proach for high-quality TSE. We leverage recent advancements in flow matching to condition our model on the target speaker’s audio. This simple yet effective approach reduces computa- tional overhead and eliminates the need for pretrained models in complex pipelines. Through extensive e...

    work page

  7. [6]

    wtimit2mix: A cock- tail party mixtures database to study target speaker extraction for normal and whispered speech,

    M. Borsdorf, Z. Pan, H. Li, and T. Schultz, “wtimit2mix: A cock- tail party mixtures database to study target speaker extraction for normal and whispered speech,” in Proc. Interspeech 2024, 2024, pp. 5038–5042

    work page 2024

  8. [7]

    Ddtse: Discriminative diffusion model for target speech extraction,

    L. Zhang, Y . Qian, L. Yu, H. Wang, H. Yang, S. Liu, L. Zhou, and Y . Qian, “Ddtse: Discriminative diffusion model for target speech extraction,” in 2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 294–301

    work page 2024

Show all 40 references
  1. [8]

    Target speaker asr with whisper,

    A. Polok, D. Klement, M. Wiesner, S. Khudanpur, J. ˇCernock`y, and L. Burget, “Target speaker asr with whisper,” arXiv preprint arXiv:2409.09543, 2024

    2024 arXiv

  2. [9]

    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 transactions on audio, speech, and language processing, vol. 26, no. 10, pp. 1702–1726, 2018

    2018

  3. [10]

    End- to-end waveform utterance enhancement for direct evaluation metrics optimization by fully convolutional neural networks,

    S.-W. Fu, T.-W. Wang, Y . Tsao, X. Lu, and H. Kawai, “End- to-end waveform utterance enhancement for direct evaluation metrics optimization by fully convolutional neural networks,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 26, no. 9, pp. 1570–1584, 2018

    2018

  4. [11]

    Generation-based target speech extraction with speech dis- cretization and vocoder,

    L. Yu, W. Zhang, C. Du, L. Zhang, Z. Liang, and Y . Qian, “Generation-based target speech extraction with speech dis- cretization and vocoder,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2024, pp. 12 612–12 616

    2024

  5. [13]

    Storm: A diffusion-based stochastic regeneration model for speech en- hancement and dereverberation,

    J.-M. Lemercier, J. Richter, S. Welker, and T. Gerkmann, “Storm: A diffusion-based stochastic regeneration model for speech en- hancement and dereverberation,”IEEE/ACM Transactions on Au- dio, Speech, and Language Processing, 2023

    2023

  6. [14]

    Analysing diffusion-based generative approaches versus discriminative approaches for speech restoration,

    ——, “Analysing diffusion-based generative approaches versus discriminative approaches for speech restoration,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023

    2023

  7. [15]

    Conditional diffusion probabilistic model for speech enhancement,

    Y .-J. Lu, Z.-Q. Wang, S. Watanabe, A. Richard, C. Yu, and Y . Tsao, “Conditional diffusion probabilistic model for speech enhancement,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 7402–7406

    2022

  8. [16]

    Tselm: Target speaker extrac- tion using discrete tokens and language models,

    B. Tang, B. Zeng, and M. Li, “Tselm: Target speaker extrac- tion using discrete tokens and language models,” arXiv preprint arXiv:2409.07841, 2024

    2024 arXiv

  9. [17]

    Enhancing intelligibility for generative target speech extraction via joint optimization with target speaker asr,

    H. Ma, R. Chen, R. Jing, X.-L. Zhang, J. Liu, and X. Li, “Enhancing intelligibility for generative target speech extraction via joint optimization with target speaker asr,” arXiv preprint arXiv:2501.14477, 2025

    2025 arXiv

  10. [18]

    Target speech extraction with conditional diffusion model,

    N. Kamo, M. Delcroix, and T. Nakatani, “Target speech extraction with conditional diffusion model,” arXiv preprint arXiv:2308.03987, 2023

    2023 arXiv

  11. [19]

    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

    2022 arXiv

  12. [20]

    Guided conditioning with predic- tive network on score-based diffusion model for speech enhance- ment,

    D. Kim, D.-H. Yang, D. Kim, J.-H. Chang, J. Yang, J. Choi, M. Lee, and H.-g. Moon, “Guided conditioning with predic- tive network on score-based diffusion model for speech enhance- ment,” in Proc. Interspeech 2024, 2024, pp. 1190–1194

    2024

  13. [21]

    V oice- filter: Targeted voice separation by speaker-conditioned spectro- gram masking,

    Q. Wang, H. Muckenhirn, K. Wilson, P. Sridhar, Z. Wu, J. Her- shey, R. A. Saurous, R. J. Weiss, Y . Jia, and I. L. Moreno, “V oice- filter: Targeted voice separation by speaker-conditioned spectro- gram masking,” arXiv preprint arXiv:1810.04826, 2018

    arXiv

  14. [22]

    Spex+: A complete time domain speaker extraction network,

    M. Ge, C. Xu, L. Wang, E. S. Chng, J. Dang, and H. Li, “Spex+: A complete time domain speaker extraction network,” arXiv preprint arXiv:2005.04686, 2020

    2005 arXiv

  15. [23]

    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

    2023 arXiv

  16. [24]

    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,” arXiv preprint arXiv:2410.06885, 2024

    2024 arXiv

  17. [25]

    Scalable diffusion models with transform- ers,

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

    2023

  18. [26]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

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

    2021

  19. [27]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” in Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , 2022, pp. 11 976–11 986

    2022

  20. [28]

    Separate and diffuse: Using a pretrained diffusion model for better source separation,

    S. Lutati, E. Nachmani, and L. Wolf, “Separate and diffuse: Using a pretrained diffusion model for better source separation,” in The Twelfth International Conference on Learning Representations , 2024

    2024

  21. [29]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  22. [30]

    Wham!: Extending speech separation to noisy environments,

    G. Wichern, J. Antognini, M. Flynn, L. R. Zhu, E. McQuinn, D. Crow, E. Manilow, and J. Le Roux, “Wham!: Extending speech separation to noisy environments,” in Proc. Interspeech, Sep. 2019

    2019

  23. [31]

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

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

    2020

  24. [32]

    Diffusion-based generative speech source separation,

    R. Scheibler, Y . Ji, S.-W. Chung, J. Byun, S. Choe, and M.-S. Choi, “Diffusion-based generative speech source separation,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  25. [33]

    Dpccn: Densely- connected pyramid complex convolutional network for robust speech separation and extraction,

    J. Han, Y . Long, L. Burget, and J. ˇCernock`y, “Dpccn: Densely- connected pyramid complex convolutional network for robust speech separation and extraction,” in IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 7292–7296

    2022

  26. [34]

    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 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1– 2

    2023

  27. [35]

    Score-based generative modeling through stochas- tic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochas- tic differential equations,” arXiv preprint arXiv:2011.13456 , 2020

    2011 arXiv

  28. [36]

    Diffusion-based speech enhancement with a weighted generative-supervised learning loss,

    J.-E. Ayilo, M. Sadeghi, and R. Serizel, “Diffusion-based speech enhancement with a weighted generative-supervised learning loss,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 12 506–12 510

    2024

  29. [37]

    Dccrn: Deep complex convolution re- current network for phase-aware speech enhancement,

    Y . Hu, Y . Liu, S. Lv, M. Xing, S. Zhang, Y . Fu, J. Wu, B. Zhang, and L. Xie, “Dccrn: Deep complex convolution re- current network for phase-aware speech enhancement,” arXiv preprint arXiv:2008.00264, 2020

    2008 arXiv

  30. [38]

    Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,

    A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hekstra, “Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,” in IEEE international conference on acoustics, speech, and signal processing (ICASSP), vol...

    2001

  31. [39]

    A short- time objective intelligibility measure for time-frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “A short- time objective intelligibility measure for time-frequency weighted noisy speech,” in IEEE international conference on acoustics, speech and signal processing (ICASSP), 2010, pp. 4214–4217

    2010

  32. [40]

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

    C. K. Reddy, V . Gopal, and R. Cutler, “Dnsmos p. 835: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2022, pp. 886–890

    2022

Pith tools

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