{"id":"f0c41f97-a28a-46e5-baa6-5755d5573ff6","arxiv_id":"2607.08111","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"Proxy-supervised joint fine-tuning of a BSRNN separator with ASR, speaker-similarity, VAD and DNSMOS losses on a new 71k real-conversation corpus yields the best SIM and timing F1 on REAL-T.","lead":"PS4 trains target-speaker extractors on real multi-talker conversations by replacing missing clean references with four proxy losses (ASR, speaker similarity, VAD, perceptual quality) and a new 71k-sample corpus. It places second on the REAL-T leaderboard with the best speaker similarity and timing F1, showing a practical path past simulated-mixture training.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Absence of loss ablations leaves unverified whether all four proxies (esp. Whisper CE and energy VAD) contribute complementary gradients that steer extraction, despite solid leaderboard ranking on SIM/F1.","rationale":"The reader correctly isolates the untested joint-proxy assumption (Eqs. 2–5) as the weakest link and already conditions the verdict on the missing ablations and λ values. The leaderboard numbers themselves are externally verifiable and the method is reproducible, so the empirical ranking claim stands; the only soft spot is the causal attribution to “four complementary” objectives. No stronger internal inconsistency or data-leakage problem is evident (DipCo, unseen in training, still improves), therefore the CONDITIONAL verdict and high confidence require no adjustment.","tokens_in":8107,"tokens_out":521,"duration_ms":36514,"concrete_test":"Re-train the BSRNN separator four times on REAL-PS4, each time dropping exactly one of L_CE, L_SIM, L_VAD, L_DNSMOS (keeping the other three and the same schedule/checkpoint), then re-evaluate TER/SIM/DNSMOS/F1 on the REAL-T development set (or a fixed 500-sample subset). If removing L_CE or L_VAD changes SIM or F1 by <0.01, the claim that all four proxies supply complementary effective signals is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim rests on PS4’s 2nd-place REAL-T ranking (best SIM 0.565 and timing F1 0.871) as evidence that the four proxy losses in Eq. (1) (Whisper CE, ResNet34 hinge ranking, energy-based VAD BCE, DNSMOS OVRL) supply effective training signals without clean references. Yet §III–IV report only the joint model; no leave-one-out, weight-sweep, or single-loss baselines appear, nor are the λ values given. Because the four training objectives are near-isomorphic to the four evaluation metrics themselves, the observed gains could be driven by only the two terms (SIM, DNSMOS) that most directly match the metrics on which PS4 wins, while CE and crude energy VAD contribute little or even noise. Without that decomposition the “complementary” and “jointly effective” assertions remain untested assumptions rather than demonstrated facts.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper introduces PS4, a proxy-supervised joint-training framework for target speaker extraction (TSE) on real conversational mixtures where clean target references are unavailable. The authors construct REAL-PS4, a 71,771-sample corpus reformatted from AISHELL-4, AliMeeting, AMI and CHiME-6, each sample providing a mixture, enrollment, transcript and frame-level VAD labels. Starting from a public BSRNN-ECAPA checkpoint, only the BSRNN separator is fine-tuned under a weighted sum of four frozen-proxy losses (Whisper large-v3 cross-entropy, ResNet34 cosine-margin ranking, energy-based VAD BCE, and DNSMOS OVRL). On the REAL-T development set and official leaderboard, PS4 outperforms the two official BSRNN baselines across TER, SIM, DNSMOS and timing F1, ranking 2nd overall and achieving the best SIM (0.565) and F1 (0.871).","tokens_in":8401,"tokens_out":1059,"duration_ms":16232,"significance":"The work addresses a genuine and practically important gap: conventional SI-SNR-style TSE training cannot be applied to real multi-talker recordings. Releasing both the REAL-PS4 corpus and training code, together with an external leaderboard evaluation that uses fixed public pre-trained models, constitutes a reproducible and useful contribution. The empirical demonstration that a multi-proxy objective can improve speaker similarity and timing consistency without clean references is of clear interest to the speech-separation and conversational-ASR communities. The result is engineering-oriented rather than theoretically novel, yet the combination of scale, public resources and competitive leaderboard placement makes the paper a solid incremental advance if the complementary-proxy claim can be substantiated.","major_comments":[{"comment":"§III, Eq. (1) and the four individual losses (Eqs. 2–5): the central methodological claim is that the four proxy objectives are complementary and jointly effective. The manuscript reports only the fully joint model; no leave-one-out, single-loss, or weight-sweep ablations appear in §IV or the tables. Because the training objectives are nearly isomorphic to the four evaluation metrics, it remains untested whether Whisper CE and energy-based VAD actually contribute useful gradients or whether the observed SIM/F1 gains are driven primarily by the ranking and DNSMOS terms. A minimal ablation table is required to support the “complementary” assertion.","section":null},{"comment":"§III, Eq. (1): the four scalar weights λ_ce, λ_sim, λ_dns, λ_vad and the ranking margin m are free parameters that are never reported. Without these values (or a statement that they were set to 1 / selected by grid search on a held-out split), the joint objective cannot be reproduced and the relative importance of each proxy remains opaque.","section":null},{"comment":"§III, Eq. (4): temporal supervision is realized by binary cross-entropy between a simple frame-energy sigmoid and the diarization-derived VAD labels. Energy is a crude activity detector that can be confounded by residual interferers or noise; the paper does not verify that this proxy actually improves timing F1 relative to a model trained without LVAD. Given that timing F1 is one of the two metrics on which PS4 claims superiority, a controlled comparison is needed.","section":null}],"minor_comments":[{"comment":"Table II caption and surrounding text: the leaderboard metric is DNSMOS-P808 while development results use DNSMOS OVRL; a short clarifying sentence would prevent reader confusion.","section":null},{"comment":"Fig. 2: the y-axis ranges differ across panels and the baseline bars for DNSMOS are near zero, making visual comparison difficult; consider a common scale or a relative-improvement plot.","section":null},{"comment":"§II: the quality-filter thresholds (target >20 % duration, ≥2 valid characters, 30 s truncation) are reasonable but their sensitivity is not discussed; a one-sentence note on how many samples were discarded would help.","section":null},{"comment":"References [19] and the REAL-T challenge URL appear twice with slightly different formatting; unify the citation style.","section":null},{"comment":"Abstract and §I: “71,771 training samples” is stated without a per-source breakdown; adding a small table or parenthetical counts would improve transparency.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent systems contribution that fits a speech-processing venue. The missing ablations are the main obstacle to acceptance; once supplied, the work should be publishable. Novelty is primarily in the engineering assembly of existing public models and corpora rather than a new algorithmic principle, which is acceptable for an applications-oriented track but may be viewed as incremental by more theoretical reviewers."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that they built a usable training set (REAL-PS4, ~72k real overlapping mixtures with enrollments, transcripts and VAD labels from AISHELL-4/AliMeeting/AMI/CHiME-6) and then fine-tuned only the BSRNN separator of a public checkpoint with four frozen proxy losses—Whisper CE, ResNet speaker ranking, energy VAD BCE, DNSMOS OVRL. On the external REAL-T leaderboard it lands 2nd overall and takes the best SIM (0.565) and timing F1 (0.871). That is a concrete, reproducible step past pure simulated-mixture training.\n\nWhat is new is the corpus construction pipeline (single-speaker enrollment extraction, overlap merging, quality filters) and the joint proxy objective itself. The individual losses already exist; packaging them for real conversational TSE and shipping both data and code is the engineering contribution. The per-dataset tables are clean: PS4 beats the two official BSRNN baselines on every sub-corpus and every metric, with the biggest lift on perceptual quality. Citations are standard and the evaluation protocol is fixed and external, so the ranking claim is solid.\n\nThe soft spot is exactly the one the stress-test flags: no leave-one-out, no weight sweep, no single-proxy baselines, and the λ values are never stated. Because the four training terms line up almost one-to-one with the four evaluation metrics, we cannot yet tell whether Whisper CE and crude energy VAD are doing real work or just riding along. That is a real but non-fatal gap; the overall system still works and the public artifacts let anyone check. Free parameters exist, but they are ordinary loss weights, not hidden fitting knobs.\n\nThis is for people who actually train meeting-style TSE or personalized ASR and need real-data recipes. The math is elementary, the data construction is careful, and the thinking is clear. I would send it to peer review; the missing ablations are fixable and the empirical result is already useful. If you work in the area, grab the corpus and try the recipe; if not, the paper is still a clean example of proxy supervision done right.","headline":"Practical proxy recipe plus a real 72k-sample corpus that actually moves the REAL-T needle, but the four-loss story is still un-ablated.","tokens_in":8971,"tokens_out":550,"would_cite":true,"duration_ms":14849,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Proxy losses from transcripts, enrollment, activity labels and quality scores can train a target-speaker extractor without any clean target audio.","keywords":["target speaker extraction","proxy supervision","joint training","real conversational speech","BSRNN","REAL-T"],"falsifier":"An ablation that removes any one of the four proxy terms (or replaces the joint objective with a conventional SI-SNR loss on simulated clean targets) and measures a statistically significant drop in speaker similarity or timing F1 on the held-out REAL-T validation set.","tokens_in":9048,"feed_emoji":"🎧","tokens_out":571,"duration_ms":18270,"temperature":0.7,"pith_summary":"Target speaker extraction needs clean reference speech for ordinary training losses, yet real conversational recordings never supply it. This paper shows that four proxy signals already present in public meeting data—ground-truth transcripts, enrollment clips, diarization-derived activity labels, and a differentiable perceptual-quality score—can jointly supervise a separator network. The authors assemble 71 771 such samples from four existing corpora and fine-tune only the separator of a pretrained BSRNN model. On the REAL-T real-mixture benchmark the resulting system ranks second overall and records the best speaker-similarity and timing-F1 scores among all entries, indicating that the proxy objectives alone are enough to steer extraction toward the intended speaker under realistic overlap, noise and reverberation.","feed_headline":"Four proxy losses train speaker extraction without clean audio","feed_subtitle":"Transcripts, enrollment clips, activity labels and quality scores replace missing clean targets on real mixtures.","key_machinery":"The joint proxy objective L = λ_ce L_CE + λ_sim L_SIM + λ_dns L_DNSMOS + λ_vad L_VAD, which back-propagates through frozen ASR, speaker-embedding and quality models into the trainable separator.","core_discovery":"Four complementary, fully differentiable proxy losses—Whisper ASR cross-entropy, enrollment-based speaker-similarity ranking, frame-level voice-activity binary cross-entropy, and DNSMOS overall quality—suffice to fine-tune a BSRNN separator for target-speaker extraction on real conversational mixtures, without any clean target waveform.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Four proxies train TSE on real mixtures without clean targets","Proxy losses fine-tune BSRNN separator for real conversational extraction","ASR CE, sim, VAD and DNSMOS replace clean audio in TSE training","Joint proxy supervision enables target extraction from real speech mixes","PS4 fine-tunes BSRNN with four proxies for clean-free real TSE"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The four proxy losses, none of which ever sees a clean target waveform, still produce gradients that correctly steer the separator toward the intended speaker rather than toward some other acoustically plausible mixture component.","fun_headline_variants_meta":{"raw":{"variants":["Four proxies train TSE on real mixtures without clean targets","Proxy losses fine-tune BSRNN separator for real conversational extraction","ASR CE, sim, VAD and DNSMOS replace clean audio in TSE training","Joint proxy supervision enables target extraction from real speech mixes","PS4 fine-tunes BSRNN with four proxies for clean-free real TSE"]},"model":"grok-4.5","effort":"low","cost_usd":0.0061,"raw_usage":{"total_tokens":1564,"prompt_tokens":725,"num_sources_used":0,"completion_tokens":95,"cost_in_usd_ticks":61000000,"prompt_tokens_details":{"text_tokens":725,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":744,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":725,"tokens_out":95,"duration_ms":6743,"temperature":1.0,"reasoning_tokens":744,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T12:52:33.240983+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"An ablation that removes any one of the four proxy terms (or replaces the joint objective with a conventional SI-SNR loss on simulated clean targets) and measures a statistically significant drop in speaker similarity or timing F1 on the held-out REAL-T validation set.","supporting_citations":[],"review_version":1}