{"id":"0daad1ac-d029-4749-8d1d-a8b3d24c70bb","arxiv_id":"2606.23332","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Introduces own-voice cancellation as a complement to target speaker extraction and benchmarks lightweight 2 ms latency models for far-field speech enhancement.","lead":"The paper introduces own-voice cancellation to remove an enrolled speaker from noisy audio mixtures while keeping other speech, using low-latency neural models conditioned on a short enrollment clip. A smart generalist might read it to understand practical fixes for audio feedback distortion in voice devices and assistants.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Conditioning on a short enrollment utterance may fail to isolate only the target speaker without suppressing other speech in the mixture.","rationale":"The reader’s weakest assumption directly identifies the same conditioning/isolation step as the load-bearing point. Because the abstract supplies no speaker-specific isolation metrics and the full text is referenced but not reproduced here, the concern remains the primary unverified link in the argument; confirming selective cancellation would move the verdict from UNVERDICTED to CONDITIONAL.","tokens_in":1670,"tokens_out":386,"duration_ms":24800,"concrete_test":"On a held-out test set containing at least three concurrent speakers (one enrolled with a 3–5 s utterance), compute (a) the change in SI-SDR for the non-target speakers between input mixture and OVC output and (b) the dB reduction in the enrolled speaker’s energy (via oracle separation). If non-target SI-SDR drops by more than 2 dB on average while target energy drops by less than 8 dB, the isolation claim does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the enrollment-conditioned model removes the enrolled speaker while leaving all other speech intact. This hinges on the auxiliary network (or its linear RNN replacement) producing a sufficiently discriminative embedding from a short utterance that the masker can null only the target without collateral suppression. In far-field noisy conditions with overlapping speakers, short-utterance embeddings are known to have high variance; if the model uses a similarity-based or embedding-concatenation mechanism (as in TD-SpeakerBeam), any overlap in embedding space would either leak the target or attenuate non-targets. The abstract reports SDR/MOS gains but does not indicate separate metrics (e.g., target-speaker energy reduction vs. non-target SI-SDR) that would confirm selective cancellation.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces own-voice cancellation (OVC) as the complement of target-speaker extraction: conditioning a 2 ms latency time-domain masker on a short enrollment utterance to remove only the enrolled speaker from a noisy multi-speaker mixture while preserving all other speech. It benchmarks TD-SpeakerBeam against a lighter Mamba-MinGRU architecture, shows that swapping the auxiliary network for a linear RNN encoder improves SDR and predicted MOS while lowering compute, and concludes that OVC is a practical objective for far-field devices whose round-trip latency would otherwise distort the user's own voice.","tokens_in":1806,"tokens_out":512,"duration_ms":13993,"significance":"If the selective-cancellation claim is substantiated, the work supplies a concrete, low-latency formulation that directly addresses a perceptual artifact in streaming far-field enhancement. The replacement of the ConvTasNet auxiliary network by a linear RNN and the introduction of the Mamba-MinGRU masker are concrete engineering contributions that could be adopted in resource-constrained devices.","major_comments":[{"comment":"Abstract and §3 (model description): the central claim requires that the enrollment-conditioned masker nulls only the target speaker without attenuating non-target speech. No separate quantitative evidence (target-speaker energy reduction, non-target SI-SDR, or speaker-specific PESQ) is referenced; aggregate SDR/MOS gains alone do not confirm selectivity, especially given known variance of short-utterance embeddings in far-field overlap.","section":"Abstract, §3"},{"comment":"§4 (experiments): the manuscript reports SDR and MOS improvements after replacing the auxiliary network, yet supplies neither dataset statistics (number of speakers, enrollment length distribution, SNR range) nor ablation isolating the contribution of the linear RNN versus the Mamba blocks. Without these, the performance delta cannot be attributed to the claimed architectural change.","section":"§4"}],"minor_comments":[{"comment":"Notation for the enrollment embedding and its concatenation or similarity mechanism with the masker input should be defined explicitly (e.g., an equation in §3) rather than left to the TD-SpeakerBeam reference.","section":"§3"},{"comment":"Figure captions and axis labels should state the exact enrollment duration used at test time and whether enrollment is clean or noisy.","section":"Figures"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. We address each major comment below.","responses":[{"response":"We agree that demonstrating selectivity is essential to substantiate the OVC claim. While the aggregate SDR and MOS results are consistent with selective cancellation of the enrolled speaker, they do not directly isolate target-speaker attenuation from non-target preservation. In the revised manuscript we will add target-speaker energy reduction and non-target SI-SDR metrics to provide explicit quantitative support for the selectivity of the enrollment-conditioned masker.","revision_made":"yes","referee_comment":"[Abstract, §3] Abstract and §3 (model description): the central claim requires that the enrollment-conditioned masker nulls only the target speaker without attenuating non-target speech. No separate quantitative evidence (target-speaker energy reduction, non-target SI-SDR, or speaker-specific PESQ) is referenced; aggregate SDR/MOS gains alone do not confirm selectivity, especially given known variance of short-utterance embeddings in far-field overlap."},{"response":"We accept that the experimental reporting is incomplete. The revised version will include full dataset statistics (speaker count, enrollment length distribution, SNR ranges) and a dedicated ablation that isolates the linear RNN encoder from the original ConvTasNet auxiliary network as well as the contribution of the Mamba blocks.","revision_made":"yes","referee_comment":"[§4] §4 (experiments): the manuscript reports SDR and MOS improvements after replacing the auxiliary network, yet supplies neither dataset statistics (number of speakers, enrollment length distribution, SNR range) nor ablation isolating the contribution of the linear RNN versus the Mamba blocks. Without these, the performance delta cannot be attributed to the claimed architectural change."}],"tokens_in":1337,"tokens_out":378,"duration_ms":20312,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that the authors treat own-voice cancellation as its own objective—removing an enrolled speaker from a noisy multi-speaker mix while keeping everything else—and build a 2 ms latency time-domain model around it. They benchmark TD-SpeakerBeam against a Mamba-MinGRU masker and report that swapping the auxiliary network for a linear RNN encoder improves both signal quality and efficiency.\n\nThe practical framing is useful. Far-field devices often stream enhanced audio back to the user, and the round-trip delay creates audible distortion on the speaker's own voice. Positioning OVC as the complement to target speaker extraction gives a clear way to think about the requirement. The architecture choices—Mamba blocks with MinGRU mixing and the lighter auxiliary—look like reasonable moves for keeping the model lightweight enough for real hardware.\n\nThe soft spot is the lack of visible numbers. The abstract states that the RNN replacement improves SDR and predicted MOS, but supplies no values, no dataset details, no error bars, and no separate metrics showing target suppression versus non-target preservation. Without those, it is difficult to judge whether the selectivity claim holds in overlapping far-field conditions. The concern about short enrollment embeddings having high variance is worth checking directly in the results; overall SDR alone would not confirm it.\n\nThis work is aimed at engineers building low-latency speech pipelines for consumer devices. A reader who needs concrete latency and compute numbers for edge deployment would find the model variants worth examining.\n\nI would send it for peer review so the quantitative claims and selectivity metrics can be examined in detail.","headline":"This paper defines own-voice cancellation as a distinct task and shows a Mamba-MinGRU model plus RNN auxiliary swap that cuts latency and compute while claiming SDR/MOS gains.","tokens_in":2309,"tokens_out":400,"would_cite":false,"duration_ms":26483,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Own-voice cancellation removes an enrolled speaker from far-field mixtures at 2 ms latency while keeping other speech.","keywords":["own-voice cancellation","far-field speech enhancement","low-latency processing","speaker enrollment","time-domain model","multi-speaker mixture","Mamba architecture"],"falsifier":"A controlled test in which the model either removes non-enrolled speakers or leaves the enrolled speaker audible after processing.","tokens_in":2601,"feed_emoji":"🔇","tokens_out":620,"duration_ms":15171,"temperature":0.7,"pith_summary":"The paper defines own-voice cancellation as the task of removing only an enrolled target speaker from a noisy multi-speaker mixture. This addresses distortion that occurs when far-field devices send enhanced audio back to the user, because round-trip delay exceeds the threshold for perceiving one's own voice as natural. A time-domain model is conditioned on a short enrollment utterance to identify the target speaker. The authors compare TD-SpeakerBeam to a lighter Mamba-MinGRU masker and show that swapping the auxiliary network for a linear RNN encoder raises signal-to-distortion ratio and predicted MOS while cutting compute. If correct, the result supplies a concrete low-latency objective for real-time far-field denoising.","feed_headline":"Cancel enrolled speaker from mixtures at 2 ms latency","feed_subtitle":"Own-voice cancellation removes only the target speaker while preserving others, fixing distortion from device feedback delay.","key_machinery":"Conditioning a time-domain masker on a short enrollment utterance to isolate and cancel only the enrolled speaker.","core_discovery":"Own-voice cancellation removes a target enrolled speaker from a noisy multi-speaker mixture while preserving any remaining speech. Framed as the complement of target speaker extraction, the task is solved by conditioning a time-domain model on a short enrollment utterance. The model achieves only 2 ms algorithmic latency, and replacing the ConvTasNet-based auxiliary network with a linear RNN encoder improves both signal-to-distortion ratio and predicted MOS while reducing compute.","pith_inferences":["Devices could stream processed audio without the user's voice creating an unnatural echo.","The same conditioning technique might extend to selective enhancement of only non-enrolled voices.","Real-time implementation on embedded hardware would confirm whether the reported 2 ms latency holds under actual acoustic conditions."],"forward_implications":["OVC functions as a practical low-latency enhancement objective for far-field denoising.","The Mamba-MinGRU masker offers a lighter alternative that matches or exceeds TD-SpeakerBeam performance.","Linear RNN encoder replacement improves both objective metrics and perceptual quality scores.","The approach directly mitigates own-voice artifacts caused by device round-trip latency."],"fun_headline_variants":["Enrolled speaker cancellation at 2 ms latency","2 ms latency model removes target speaker from mixtures","Lightweight RNN encoder for 2 ms own-voice cancellation","Own-voice cancellation conditioned on enrollment at 2 ms"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Conditioning on a short enrollment utterance suffices to isolate and remove only the target speaker without harming other speech.","fun_headline_variants_meta":{"raw":{"variants":["Enrolled speaker cancellation at 2 ms latency","2 ms latency model removes target speaker from mixtures","Lightweight RNN encoder for 2 ms own-voice cancellation","Own-voice cancellation conditioned on enrollment at 2 ms"]},"model":"grok-4.3","cost_usd":0.007841,"raw_usage":{"total_tokens":3551,"prompt_tokens":614,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":78412000,"prompt_tokens_details":{"text_tokens":614,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2882,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":614,"tokens_out":55,"duration_ms":23832,"temperature":1.0,"reasoning_tokens":2882,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T06:47:34.460400+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A controlled test in which the model either removes non-enrolled speakers or leaves the enrolled speaker audible after processing.","supporting_citations":[],"review_version":1}