Breaking the Quality--Intelligibility Trade-off in Streaming Target Speaker Extraction via Deep-Feature-Anchored Preference Optimization

Guiping Zhong; Hui Lu; Jiakui Li; Jinjiang Liu; Liyang Chen; Shiyin Kang; Shuhai Peng; Zhiyong Wu

REVIEW 3 major objections 5 minor 16 references

The quality–intelligibility trade-off in streaming target-speaker extraction is caused by a bad optimization anchor, not by streaming itself; WavLM-ranked DPO plus a larger Conformer kernel breaks it with a 10.9% WER gain and simultaneous q

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 · grok-4.5

2026-07-14 13:37 UTC pith:UH3QFPIX

load-bearing objection Solid engineering paper: the trade-off is mostly the preference anchor, not streaming itself, and WavLM-ranked DPO plus a larger Conformer kernel gives a clean multi-metric win under 560 ms latency. the 3 major comments →

arxiv 2607.10191 v1 pith:UH3QFPIX submitted 2026-07-11 cs.SD cs.AI

Breaking the Quality--Intelligibility Trade-off in Streaming Target Speaker Extraction via Deep-Feature-Anchored Preference Optimization

Shuhai Peng , Jinjiang Liu , Hui Lu , Liyang Chen , Guiping Zhong , Jiakui Li , Shiyin Kang , Zhiyong Wu This is my paper

classification cs.SD cs.AI

keywords Direct Preference OptimizationStreaming TSEQuality–Intelligibility Trade-offDeep Feature AnchoringReward HackingWavLMConformerTarget Speaker Extraction

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.

Generative streaming models for target speaker extraction usually force a choice: chase perceptual quality and lose intelligibility, or chase word-error rate and lose naturalness. This paper argues that the trade-off is not forced by causal streaming architectures. It is caused by optimizing against the wrong anchor. Directly maximizing perceptual scores such as DNSMOS makes the model erase stop consonants and fricatives, producing smooth audio that scores well but is hard to understand. The authors show two complementary fixes. First, enlarging the Conformer convolution kernel to size 15 gives the encoder richer local spectro-temporal context under the same streaming constraint. Second, they fine-tune only the semantic pathway with Direct Preference Optimization whose preference pairs are ranked by WavLM cosine similarity to a clean reference. Because WavLM already encodes both phonetic structure and speaker identity, the ranking resists the reward-hacking that surface quality metrics invite. Under a 560 ms chunk, the combined system lowers word error rate from 0.138 to 0.123 while also improving background quality and speaker similarity.

Core claim

The quality–intelligibility trade-off observed in generative streaming target speaker extraction is not an architectural necessity of causal models; it is an artifact of an inappropriate optimization anchor. Ranking Direct Preference Optimization pairs by WavLM cosine similarity supplies a deep-feature anchor that jointly preserves phonetic content and acoustic naturalness, producing a simultaneous improvement in word error rate, perceptual quality, and speaker similarity.

What carries the argument

WavLM-anchored DPO: preference pairs generated from the same candidates are ranked by cosine similarity of WavLM embeddings to a clean reference, and the DPO loss is applied only to the semantic pathway (shared Conformer encoder and SELM) while the acoustic pathway stays frozen; together with a Conformer convolution kernel of size 15.

Load-bearing premise

That WavLM cosine similarity to a clean reference is a sufficiently complete and hard-to-hack proxy for both phonetic structure and speaker identity, so maximizing it will automatically keep the consonants and fricatives that intelligibility needs.

What would settle it

Construct preference pairs ranked by WavLM on the same 560 ms streaming setup and show that the resulting model still erases stop bursts and fricatives (visible in mel spectrograms) while WER fails to drop below the k=15 baseline, or that DNSMOS-ranked DPO can be regularized to match WavLM’s Pareto gains without the deep feature.

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

If this is right

Streaming generative TSE can be improved without sacrificing causality or real-time factor by changing only the preference ranking metric and local convolution kernel.
Surface perceptual metrics such as DNSMOS are actively harmful as sole DPO anchors for speech generation because they reward erasure of phonetic content.
Deep self-supervised speech features can serve as stable multi-objective anchors that jointly protect intelligibility and naturalness.
Restricting DPO gradients to the semantic pathway is enough to improve WER while leaving acoustic reconstruction quality intact.
The same deep-feature ranking idea is proposed by the authors as a route to lower-latency (<200 ms) and multi-speaker regimes.

Where Pith is reading between the lines

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

Any proxy that scores only waveform smoothness is likely to produce the same consonant-erasure failure mode in other generative speech tasks, not only TSE.
The success of WavLM ranking suggests that other large self-supervised speech models could serve as interchangeable anchors, provided they jointly encode phonetics and speaker identity.
Freezing the acoustic pathway during DPO may be a general pattern for keeping codec-based generators from drifting when only discrete semantic tokens control intelligibility.
If the method scales to noisier multi-talker mixtures, the same ranking could reduce reliance on separate multi-objective reward models that are themselves hackable.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Solid engineering paper: the trade-off is mostly the preference anchor, not streaming itself, and WavLM-ranked DPO plus a larger Conformer kernel gives a clean multi-metric win under 560 ms latency.

read the letter

The useful takeaway is simple. In generative streaming TSE the quality–intelligibility trade-off is largely an optimization-anchor problem, not an architectural inevitability. When they rank the same 16 stochastic candidates three different ways, DNSMOS ranking produces classic reward hacking (smooth spectrograms that erase stop bursts and fricatives, WER collapse), WER ranking is too narrow and hurts quality, and WavLM cosine similarity to a clean reference yields a Pareto step: WER 0.138 → 0.123 (kernel enlargement does half the work, DPO the rest) with simultaneous small gains in BAK/OVL and speaker similarity. That controlled comparison is the real contribution.

What they do well: the diagnosis is concrete (mel spectrograms in Fig. 3 make the phonetic erasure visible), the ablations are clean (identical candidate pool, full β×epoch sweep in the appendix, kernel-size table showing k=3 hurts and k=15 helps), and they freeze the acoustic pathway so DPO only touches the semantic tokens that actually drive WER. The StarTSE backbone is their own prior work, but the new pieces—kernel enlargement plus deep-feature anchoring—are clearly isolated.

Soft spots are ordinary engineering caveats, not load-bearing flaws. Everything is on Libri2Mix; no error bars; β and checkpoint are chosen post-hoc; no code release. The central premise that WavLM is hard enough to hack is supported inside their experiments but remains untested on noisier multi-speaker data or lower latency. Gains are modest. None of that overturns the internal claim.

This is for people building real-time voice interfaces or generative speech separation who already care about DPO and preference metrics. It is not a field-reorganizing result, but it is a clear, reproducible-looking lesson about how to choose the ranking signal. I would send it to peer review; the evidence is sharp enough to deserve referee time and the usual revision requests. Worth reading if you work in the area; I would cite the reward-hacking diagnosis and the three-way ranking comparison.

Referee Report

3 major / 5 minor

Summary. The paper argues that the quality–intelligibility trade-off observed in generative streaming Target Speaker Extraction (TSE) is caused by an inappropriate optimization anchor rather than by streaming architectural constraints. Building on the StarTSE backbone, the authors replace the Transformer encoder with a causal Conformer (kernel k=15) and fine-tune only the semantic pathway (Shared Encoder + SELM) via Direct Preference Optimization (DPO). Preference pairs are constructed from 16 stochastic candidates per mixture and ranked by three independent criteria: DNSMOS, WER, and WavLM cosine similarity to a clean reference. Only the WavLM ranking yields a Pareto improvement under 560 ms causal chunks: WER 0.138→0.123 (10.9 % relative), with simultaneous modest gains in DNSMOS BAK/OVL and speaker similarity. DNSMOS ranking produces classic reward hacking (spectrally smooth but phonetically erased speech, WER collapse), while WER ranking is overly conservative. Full β×epoch sweeps and a kernel-size ablation support the claim.

Significance. If the result holds, the work supplies a concrete, reproducible diagnosis of reward hacking in streaming generative TSE and a practical remedy (deep-feature-anchored DPO) that simultaneously improves the three metrics that matter for deployment. The controlled design—identical candidate pools ranked three ways, frozen acoustic pathway, full hyper-parameter tables—makes the causal attribution to the ranking criterion unusually clear for this literature. The approach is immediately usable by practitioners already employing StarTSE-style models and suggests a broader principle: preference anchors that jointly encode phonetic structure and acoustic fidelity are more robust than surface perceptual or transcription-only metrics. Strengths include the transparent ablation structure, the qualitative spectrogram evidence of the hacking mechanism, and the complete Appendix Table 4.

major comments (3)

§4.2 / Table 2 and Appendix Table 4: the headline configuration (DPOWavLM, β=0.9, e0) is selected post-hoc from a 5×3 grid. While the full grid is reported, no statistical significance, confidence intervals, or multiple-comparison correction is provided for the 0.129→0.123 WER step. Given that several nearby cells already sit at 0.129–0.134, a formal test (or at least bootstrap standard errors over the 3 000-utterance test set) is needed to establish that the claimed 4.7 % pure-DPO gain is reliable rather than a lucky checkpoint.
§2.2 and §3.1: all preference pairs and all evaluation are confined to Libri2Mix (clean two-speaker mixtures). The central claim that WavLM cosine similarity is an unhackable, sufficiently complete proxy for phonetic structure is therefore demonstrated only under matched, high-SNR conditions. Without at least one mismatched or noisy evaluation set, it remains open whether the same ranking criterion continues to resist reward hacking when the clean reference itself is imperfect or when the mixture contains non-speech interference.
§2.1 (Acoustic Pathway) and the axiom that freezing ARLM+codec is sufficient: the paper never reports an ablation that unfreezes the acoustic pathway or jointly optimizes both pathways under the same WavLM ranking. If the acoustic pathway can still introduce intelligibility-destroying artifacts after the semantic tokens are fixed, the claimed decoupling is incomplete. A short controlled experiment (or an explicit statement of why it is unnecessary) would close this load-bearing assumption.

minor comments (5)

Table 1: the Proposed row reports DNSMOS SIG 3.491, which is lower than both StarTSE (3.535) and the k=15 baseline (3.541). The text claims “simultaneous gains in audio quality”; this should be qualified to BAK/OVL only, or the SIG drop should be discussed.
Fig. 3 caption and §4.3: the spectrogram comparison is persuasive but only one utterance is shown. Adding one or two additional examples (or a quantitative measure of high-frequency energy preservation) would strengthen the visual claim.
§2.2: the number of stochastic candidates (16) and the precise sampling temperature / top-k settings used to generate them are not stated; these details affect reproducibility of the preference pairs.
References: StarTSE is cited as arXiv:2604.19635; ensure the final published version (if available) is used, and that the relationship between the two papers is made explicit for readers who encounter only one of them.
Notation: the DPO loss (Eq. 1) uses πθ(yw|x) while the model actually generates discrete semantic tokens U; a short clarification that the likelihood is taken over the SELM token sequence would remove ambiguity.

Circularity Check

0 steps flagged

No significant circularity: empirical DPO ranking comparisons on external metrics; minor self-citation of StarTSE backbone is not load-bearing for the Pareto claim.

full rationale

The paper's central claim (quality–intelligibility trade-off is anchor-driven, broken by WavLM-ranked DPO plus k=15 kernel) is established by controlled experiments: identical candidate pools from the same mixtures are ranked three ways (DNSMOS / WER / WavLM), DPO is run with the standard loss (Eq. 1), and results are measured by independent external evaluators (Whisper WER, DNSMOS, WavLM/WeSpeaker cosine). No equation or definition reduces the reported WER drop (0.138 o0.123) or simultaneous BAK/OVL/Sim gains to a quantity defined by the optimized metric itself. Preference construction (Fig. 2) and the three-variant ablation (Table 2, full sweep Table 4) are falsifiable against public baselines and the authors' own prior checkpoint. Self-citation of StarTSE [6] supplies only the streaming backbone; the novel contribution (anchor choice) is independently demonstrated and does not rely on an unverified uniqueness theorem or fitted parameter renamed as prediction. Score 1 reflects only the ordinary presence of same-group prior work, which is not circular under the stated criteria.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard DPO mathematics, the empirical properties of a pre-trained WavLM encoder, the causal streaming protocol of StarTSE, and a handful of hand-chosen hyper-parameters (kernel size, β, number of candidates). No new physical entities are postulated; the free parameters are ordinary ML knobs whose values are reported.

free parameters (4)

Conformer convolution kernel size k = 15
Chosen by ablation (k=3 hurts, k=15 helps); not derived from first principles.
DPO KL coefficient β = 0.9 (selected)
Swept over {0.1,0.3,0.5,0.7,0.9}; best reported result uses β=0.9.
Number of stochastic candidates per mixture = 16
Fixed at 16 for preference-pair construction; no sensitivity study.
DPO learning rate and warmup = 1e-5 / 200 steps
Adam 1e-5 with 200-step warmup; standard but still free.

axioms (4)

standard math Standard DPO loss (Eq. 1) correctly aligns a policy to pairwise preferences without an explicit reward model.
Taken from Rafailov et al. 2023; used unchanged.
domain assumption WavLM embeddings jointly encode phonetic structure and speaker identity sufficiently well that cosine similarity to a clean reference is a hacking-resistant ranking signal.
Stated in §2.2 and §4.3; never independently verified beyond the observed multi-metric gains.
ad hoc to paper Freezing the acoustic pathway (ARLM + codec) while updating only the semantic pathway is sufficient to improve intelligibility without harming reconstruction.
Design choice justified by the claim that WER is determined by discrete semantic tokens; not ablated against full end-to-end DPO.
domain assumption Whisper Large-V3 WER is a faithful proxy for human intelligibility under the evaluated conditions.
Used as the primary intelligibility metric throughout; common practice but still an external model assumption.

pith-pipeline@v1.1.0-grok45 · 15839 in / 3149 out tokens · 39240 ms · 2026-07-14T13:37:09.985139+00:00 · methodology

0 comments

read the original abstract

Generative streaming models for Target Speaker Extraction (TSE) commonly exhibit a quality--intelligibility trade-off, wherein naive optimization for perceptual audio quality tends to degrade speech intelligibility, and conversely. We reveal that this trade-off arises not from the constraints of streaming architectures, but from an inappropriate choice of optimization anchor. Directly optimizing against audio quality metrics induces catastrophic reward hacking, where content critical to pronunciation and intelligibility is systematically erased to maximize a proxy score. To break this bottleneck, we propose two complementary improvements: an enlarged Conformer convolution kernel for richer local spectro-temporal modeling, and WavLM-anchored Direct Preference Optimization (DPO) fine-tuning strategy. DPO preference pairs are ranked by WavLM cosine similarity, a deep acoustic feature encoding both phonetic structure and speaker identity, providing an optimization anchor that resists hacking. Under a 560 ms streaming chunk size, the proposed method achieves a 10.9% relative intelligibility improvement (word error rate: 0.138 to 0.123), with marginal simultaneous gains in audio quality and speaker similarity.

Figures

Figures reproduced from arXiv: 2607.10191 by Guiping Zhong, Hui Lu, Jiakui Li, Jinjiang Liu, Liyang Chen, Shiyin Kang, Shuhai Peng, Zhiyong Wu.

**Figure 1.** Figure 1: Overview of the StarTSE architecture with the proposed DPO alignment strategy. The semantic pathway (Shared Conformer Encoder and SELM, highlighted in blue) receives gradient updates during DPO, while the acoustic pathway (ARLM and Codec Decoder, shaded in gray) remains frozen. Acoustic Pathway. This pathway comprises two components: 1. Acoustic Refinement Language Model (ARLM): A language model that reco… view at source ↗

**Figure 2.** Figure 2: Preference pair construction pipeline. For each mixture, 16 candidates are generated via stochastic decoding and evaluated along three dimensions (DNSMOS, WER, WavLM). The candidates are ranked independently by each of the three criteria; the top-ranked candidate is selected as yw and the bottom-ranked as yl, producing three distinct preference pairs from the same underlying outputs. DPOWER (Intelligibili… view at source ↗

**Figure 3.** Figure 3: Mel spectrograms of the same utterance under three conditions (same sample as [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 1 linked inside Pith

[1]

IEEE Signal Processing Magazine, 40(3), 8–29 (2023) Deep-Feature-Anchored DPO for Streaming TSE 13

Zmolikova, K., Delcroix, M., Ochiai, T., Kinoshita, K., Cernocky, J., Yu, D.: Neural target speech extraction: An overview. IEEE Signal Processing Magazine, 40(3), 8–29 (2023) Deep-Feature-Anchored DPO for Streaming TSE 13

2023
[2]

In: Proc

Ge, M., Xu, C., Wang, L., Chng, E.S., Dang, J., Li, H.: SpEx+: A complete time domain speaker extraction network. In: Proc. Interspeech, pp. 1406–1410 (2020)

2020
[3]

arXiv preprint (2024)

Wang, S., Zhang, K., Lin, S., Li, J., Wang, X., Ge, M., Yu, J., Qian, Y., Li, H.: WeSep: A scalable and flexible toolkit towards generalizable target speaker extraction. arXiv preprint (2024)

2024
[4]

arXiv preprint (2024)

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

2024
[5]

In: 2025 IEEE Automatic Speech Recog- nition and Understanding Workshop (ASRU) (2025)

Tang, B., Zeng, B., Li, M.: LauraTSE: Target speaker extraction using auto- regressive decoder-only language models. In: 2025 IEEE Automatic Speech Recog- nition and Understanding Workshop (ASRU) (2025)

2025
[6]

arXiv preprint arXiv:2604.19635 (2026)

Peng, S., Lu, H., Liu, J., Chen, L., Zhong, G., Li, J., Wang, H., Li, H., Cao, L., Kang, S., Wu, Z.: StarTSE: Towards streaming target speaker extraction via chunk-wise interleaved splicing of autoregressive language model. arXiv preprint arXiv:2604.19635 (2026)

Pith/arXiv arXiv 2026
[7]

In: Advances in Neural Information Processing Systems (NeurIPS) (2023)

Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Di- rect preference optimization: Your language model is secretly a reward model. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)

2023
[8]

MIT Press, Cambridge (1998)

Sutton, R.S., Barto, A.G.: Reinforcement learning: An introduction. MIT Press, Cambridge (1998)

1998
[9]

In: Proc

Reddy, C.K.A., Gopal, V., Cutler, R.: DNSMOS: A non-intrusive perceptual ob- jective speech quality metric to measure noise suppressors. In: Proc. IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 886–890 (2021)

2021
[10]

In: Proc

Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Ro- bust speech recognition via large-scale weak supervision. In: Proc. International Conference on Machine Learning (ICML) (2023)

2023
[11]

IEEE Journal of Selected Topics in Signal Processing, 16(6), 1505–1518 (2022)

Chen,S.,Wang,C.,Chen,Z.,Wu,Y.,Liu,S.,Chen,Z.,Li,J.,Kanda,N.,Yoshioka, T., Xiao, X., et al.: WavLM: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 16(6), 1505–1518 (2022)

2022
[12]

arXiv preprint (2020)

Cosentino, J., Pariente, M., Cornell, S., Deleforge, A., Vincent, E.: LibriMix: An open-source dataset for generalizable speech separation. arXiv preprint (2020)

2020
[13]

In: Proc

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

2015
[14]

arXiv preprint (2023)

Du, Z., Zhang, S., Hu, K., Zheng, S.: FunCodec: A fundamental, reproducible and integrable open-source toolkit for neural speech codec. arXiv preprint (2023)

2023
[15]

arXiv preprint (2022)

Wang, H., Liang, C., Wang, S., Chen, Z., Zhang, B., Xiang, X., Deng, Y., Qian, Y.: WeSpeaker: A research and production oriented speaker embedding learning toolkit. arXiv preprint (2022)

2022
[16]

Springer, Berlin (2005) 14 S

Ehrgott, M.: Multicriteria optimization, 2nd edn. Springer, Berlin (2005) 14 S. Peng et al. A Full Experimental Results T able 4.CompleteDPOexperimentalresults.Allmodelsuse560msstreaminglatency. StrategyβEp. DNSMOS SIG DNSMOS BAK DNSMOS OVL WER WavLM Sim StarTSE – – 3.5350 3.7520 3.1170 0.1380 0.9590 DNSMOS-first 0.1 e0 3.5814 3.8201 3.1911 0.1609 0.9547 ...

arXiv 2005

[1] [1]

IEEE Signal Processing Magazine, 40(3), 8–29 (2023) Deep-Feature-Anchored DPO for Streaming TSE 13

Zmolikova, K., Delcroix, M., Ochiai, T., Kinoshita, K., Cernocky, J., Yu, D.: Neural target speech extraction: An overview. IEEE Signal Processing Magazine, 40(3), 8–29 (2023) Deep-Feature-Anchored DPO for Streaming TSE 13

2023

[2] [2]

In: Proc

Ge, M., Xu, C., Wang, L., Chng, E.S., Dang, J., Li, H.: SpEx+: A complete time domain speaker extraction network. In: Proc. Interspeech, pp. 1406–1410 (2020)

2020

[3] [3]

arXiv preprint (2024)

Wang, S., Zhang, K., Lin, S., Li, J., Wang, X., Ge, M., Yu, J., Qian, Y., Li, H.: WeSep: A scalable and flexible toolkit towards generalizable target speaker extraction. arXiv preprint (2024)

2024

[4] [4]

arXiv preprint (2024)

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

2024

[5] [5]

In: 2025 IEEE Automatic Speech Recog- nition and Understanding Workshop (ASRU) (2025)

Tang, B., Zeng, B., Li, M.: LauraTSE: Target speaker extraction using auto- regressive decoder-only language models. In: 2025 IEEE Automatic Speech Recog- nition and Understanding Workshop (ASRU) (2025)

2025

[6] [6]

arXiv preprint arXiv:2604.19635 (2026)

Peng, S., Lu, H., Liu, J., Chen, L., Zhong, G., Li, J., Wang, H., Li, H., Cao, L., Kang, S., Wu, Z.: StarTSE: Towards streaming target speaker extraction via chunk-wise interleaved splicing of autoregressive language model. arXiv preprint arXiv:2604.19635 (2026)

Pith/arXiv arXiv 2026

[7] [7]

In: Advances in Neural Information Processing Systems (NeurIPS) (2023)

Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Di- rect preference optimization: Your language model is secretly a reward model. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)

2023

[8] [8]

MIT Press, Cambridge (1998)

Sutton, R.S., Barto, A.G.: Reinforcement learning: An introduction. MIT Press, Cambridge (1998)

1998

[9] [9]

In: Proc

Reddy, C.K.A., Gopal, V., Cutler, R.: DNSMOS: A non-intrusive perceptual ob- jective speech quality metric to measure noise suppressors. In: Proc. IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 886–890 (2021)

2021

[10] [10]

In: Proc

Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Ro- bust speech recognition via large-scale weak supervision. In: Proc. International Conference on Machine Learning (ICML) (2023)

2023

[11] [11]

IEEE Journal of Selected Topics in Signal Processing, 16(6), 1505–1518 (2022)

Chen,S.,Wang,C.,Chen,Z.,Wu,Y.,Liu,S.,Chen,Z.,Li,J.,Kanda,N.,Yoshioka, T., Xiao, X., et al.: WavLM: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 16(6), 1505–1518 (2022)

2022

[12] [12]

arXiv preprint (2020)

Cosentino, J., Pariente, M., Cornell, S., Deleforge, A., Vincent, E.: LibriMix: An open-source dataset for generalizable speech separation. arXiv preprint (2020)

2020

[13] [13]

In: Proc

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

2015

[14] [14]

arXiv preprint (2023)

Du, Z., Zhang, S., Hu, K., Zheng, S.: FunCodec: A fundamental, reproducible and integrable open-source toolkit for neural speech codec. arXiv preprint (2023)

2023

[15] [15]

arXiv preprint (2022)

Wang, H., Liang, C., Wang, S., Chen, Z., Zhang, B., Xiang, X., Deng, Y., Qian, Y.: WeSpeaker: A research and production oriented speaker embedding learning toolkit. arXiv preprint (2022)

2022

[16] [16]

Springer, Berlin (2005) 14 S

Ehrgott, M.: Multicriteria optimization, 2nd edn. Springer, Berlin (2005) 14 S. Peng et al. A Full Experimental Results T able 4.CompleteDPOexperimentalresults.Allmodelsuse560msstreaminglatency. StrategyβEp. DNSMOS SIG DNSMOS BAK DNSMOS OVL WER WavLM Sim StarTSE – – 3.5350 3.7520 3.1170 0.1380 0.9590 DNSMOS-first 0.1 e0 3.5814 3.8201 3.1911 0.1609 0.9547 ...

arXiv 2005