{"id":"9dcfc130-ef97-4272-9eda-47a4db65ce0b","arxiv_id":"2507.02192","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"MSGLA is an iterative phase reconstruction framework that combines STFT consistency with geometric constraints; its noise-phase variant achieves small but consistent gains in background suppression metrics.","lead":"A new iterative phase estimation method for speech enhancement, called MSGLA, alternates Griffin-Lim style projections between speech and noise spectrograms to resolve the sign ambiguity of geometry-based phase reconstruction. The paper shows that on standard benchmarks the method matches or slightly beats existing phase estimators, with the largest gains in background noise suppression.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of unsupervised sign resolution rests on an unproven convergence assertion: the paper shows neither that MSGLA iterations reach one of the two geometric candidates nor, crucially, that the correct branch is selected rather than any branch.","rationale":"The reader's weakest assumption identified exactly the same gap: the convergence of the alternating projections to one of the two phase candidates is asserted from empirical observation rather than established. My independent reading confirms that this is the most load-bearing concern because the entire contribution is framed as an unsupervised resolution of the sign ambiguity in Equation (4). Without a fixed-point or basin-of-attraction analysis, the mechanism could be selecting a branch for reasons unrelated to correctness, and the oracle results would then not support the central narrative. I also note the reader's secondary concern about NP-MSGLA not directly implementing the law-of-sines derivation in Equations (8)-(9); that is a real inconsistency between the derivation and the update in Equation (12), but it is secondary because the algorithm can be judged empirically on its own. The reported metric gains are small and lack statistical validation, but that would only weaken the headline claim, not break it. The appropriate disposition remains CONDITIONAL, matching the reader's verdict, contingent on supplying the missing convergence and sign-selection evidence. I see no grounds to reject the paper outright, since the oracle experiments and reproducible framework are concrete contributions.","tokens_in":9113,"tokens_out":12511,"duration_ms":143137,"concrete_test":"Re-run the oracle experiments of Table 1 with iteration counts {1, 2, 5, 10, 20} and with three initializations (P_Y, P_Y+π, and a random phase), recording the converged phase in every T-F bin. Classify each bin as matching P_Y+|ΔP|, P_Y−|ΔP|, or neither (within, say, 0.1 rad), and compute the correct-sign fraction against ground truth. If a substantial fraction of bins fall into 'neither', or the converged branch flips with initialization, the reported convergence and sign-resolution claim is unsupported. If instead at least 95% of bins match one of the two candidates and the correct-sign fraction is robust across initializations and iteration counts, the empirical mechanism is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 states that after roughly five iterations P_X converges to either P_Y+|ΔP| or P_Y−|ΔP|, and Section 3.2 makes the analogous claim for NP-MSGLA, asserting that this 'resolves the phase ambiguity in an unsupervised manner.' This assertion is load-bearing: the paper's novelty is replacing the DNN sign predictor with an iterative sign-resolution mechanism. Yet no fixed-point analysis, stability conditions, or iteration-count sensitivity is provided. 'Either' candidate is not enough: both phase candidates are equally compatible with the measured magnitudes and noisy phase, so both are stationary points of the geometric constraint when the STFT projection is identity. In that scalar setting, a valid triple (A_X=1, A_Z=0.5, A_Y=1.2, P_Y=0) stalls at a spurious 0-phase fixed point rather than converging to either candidate at ±Δ, so geometry alone does not enforce the claimed convergence. The full STFT consistency coupling may break the tie, but the paper neither analyzes that coupling nor reports the per-bin agreement between the converged phase and the two candidates, nor the fraction of bins where the correct sign is selected. If the iteration frequently picks the wrong branch, the oracle cosine similarities (0.74–0.87) do not demonstrate successful sign resolution, and the gains over the sign predictor are not attributable to the claimed mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an iterative phase estimation framework, MSGLA, for single-channel speech enhancement under additive noise. Two variants are introduced: NM-MSGLA, which alternates Griffin-Lim-style consistency projections on speech and noise spectrograms using estimated speech and noise magnitudes, and NP-MSGLA, which uses the estimated speech magnitude and an estimated noise phase together with the law of sines to construct phase candidates. The authors report oracle experiments on VB-DMD and blind evaluations on VB-DMD and WSJ0-CHiME3, comparing against a direct phase estimator, conventional GLA, and a DNN-based sign predictor. The central claimed contribution is that the iterative procedure converges, without a trained sign classifier, to one of the two geometrically valid phase candidates, thereby resolving the sign ambiguity of Eq. (4) in an unsupervised manner.","tokens_in":9376,"tokens_out":2889,"duration_ms":38571,"significance":"If the claimed unsupervised sign resolution is correct, the paper would offer a conceptually appealing way to combine geometric phase constraints with STFT consistency, potentially reducing the need for a separate DNN sign predictor. The oracle experiments are a useful contribution because they establish upper bounds and isolate the role of each estimated input. The paper also introduces a law-of-sines formulation for phase candidate construction that, to my knowledge, is new. However, the significance is currently limited by the absence of any theoretical or empirical verification of the central convergence mechanism, and by the small, unquantified performance differences in the blind evaluations. The manuscript is honest in reporting that convergence is an empirical observation, and the comparisons are fairly presented, but the load-bearing claim remains unsupported.","major_comments":[{"comment":"The central claim of the paper, that MSGLA 'converges to either P_Y + |ΔP̂| or P_Y − |ΔP̂|' and thereby resolves the sign ambiguity, is only supported by the sentence 'Empirically, we observe...' and the analogous statement in Section 3.2. No fixed-point analysis, contraction argument, or stability condition is given, and the number of iterations is fixed at 5 with no sensitivity analysis. This is load-bearing because the entire novelty of the method rests on replacing the trained sign predictor with this iterative mechanism. As a concrete counterexample to the claim as stated, in the scalar case where the STFT projection is the identity, starting from P_X = P_Y with A_Y > A_X, A_Z, Eqs. (5)-(7) give P_Z = 0 and then P_X' = 0, so the iteration stalls at the noisy phase and does not converge to either geometric candidate. The STFT consistency coupling may prevent such stalling in practice, but the paper neither analyzes that coupling nor reports the fraction of time-frequency bins that actually converge to one of the two candidates, nor the fraction that select the correct branch. I request a proof or a precisely stated sufficient condition for convergence, together with per-bin convergence statistics over the test sets.","section":"Section 3.1, Eqs. (5)-(7) and Section 3.2, Eqs. (10)-(12)"},{"comment":"The law-of-sines formulation in Eq. (8) leads to the two candidate solutions in Eq. (9) only when |(A_Y / A_X) sin(P_Y − P_Z)| ≤ 1. In many time-frequency bins, particularly where the estimated speech magnitude is small or the phase difference is unfavorable, the argument of the inverse sine exceeds unity and the candidates do not exist. The paper does not state how such bins are handled in NP-MSGLA, nor whether this limitation affects the convergence behavior or the reported results. Since NP-MSGLA is one of the two proposed variants, this domain issue needs to be addressed explicitly, either by a fallback rule or by an analysis of how often the condition is violated on the test data.","section":"Section 3.2, Eqs. (8)-(9)"},{"comment":"The blind evaluation results show differences of only 0.01 to 0.05 in PESQ, ESTOI, SI-SNR, and CBAK between the proposed methods and the baselines, and some differences favor the baselines (e.g., NM-MSGLA PESQ 3.44 vs. GLA 3.45 on VB-DMD; NM-MSGLA SI-SNR 14.89 vs. direct phase estimator 15.88 on WSJ0-CHiME3). No confidence intervals, standard deviations, or significance tests are provided, and it is not stated how many utterances or random seeds are used. The conclusion that MSGLA 'matches well or slightly outperforms' existing algorithms, particularly in background suppression, is not statistically supported by the reported numbers. At minimum, the authors should report variance across test utterances and a paired significance test (e.g., Wilcoxon signed-rank) for the key comparisons.","section":"Tables 2 and 3"}],"minor_comments":[{"comment":"The phrase 'multi-mource Griffin-Lim algorithm' in Section 3 contains a typo; it should be 'multi-source'.","section":"Abstract and Section 1"},{"comment":"In the text, 'SI-NSR' should be 'SI-SNR', and 'NP-MSGL' should be 'NP-MSGLA'.","section":"Table 1 caption and Section 5.1"},{"comment":"The cosine similarity values in Table 1 are reported only as averages. Since phase similarity is highly variable across frequency bins and utterances, reporting a standard deviation or a per-bin histogram would help the reader judge the consistency of the proposed method.","section":"Section 5.1, Table 1"},{"comment":"The sign predictor baseline is described as a separate compact TF-GridNet model with 219K parameters. It would be helpful to state explicitly whether this model is trained jointly with the magnitude estimator or independently, and to report the sign prediction accuracy on the test sets.","section":"Section 5.2, sign predictor baseline"},{"comment":"Reference [31] is a textbook and is cited for the law of sines and cosines; a more specific citation to a trigonometric identity or a prior phase-estimation paper would be more informative. Also, reference [22] is self-cited and is not yet published; the reader would benefit from a short explanation of its consistency-preserving loss and how it relates to the present work.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and relevant problem, and the oracle experiments are a useful reference point for the community. My main concern is that the core mechanism—unsupervised sign resolution via alternating projections—is asserted without any supporting analysis, and the counterexample I constructed in the scalar setting shows that the claim is not true in general without additional conditions. I would be willing to reconsider after the authors either prove a sufficient condition for convergence, provide per-bin convergence and branch-selection statistics, or explicitly scope the claim to the setting where STFT consistency provides the needed coupling. In addition, the reported performance differences are too small to support the stated conclusions without statistical testing. The manuscript is otherwise clearly written and the experimental design is reasonable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it alternates GLA-style STFT consistency projections between the speech and noise components to resolve the geometric sign ambiguity without a trained sign predictor. The law-of-sines variant (NP-MSGLA) is also new, as far as I know, and the observation that noise phase is easier to estimate in low-energy regions is a nice empirical motivation. The oracle experiments usefully show that noise magnitude matters more than speech magnitude in this formulation.\n\nWhat goes soft: the paper's central mechanism—that MSGLA converges to one of the two phase candidates after about five iterations—is supported only by inspection. There is no fixed-point analysis, no iteration-count sensitivity, no per-bin agreement with the two candidates, and no fraction-of-bins statistic for correct branch selection. The scalar example in the stress test is enough to make the point: with magnitudes (A_X=1, A_Z=0.5, A_Y=1.2, P_Y=0), the direct update stalls at P_X=0, which is not one of the two geometric candidates. In the full STFT setting, the time-frequency consistency coupling may break that degeneracy, but the paper never shows it. This is load-bearing because the novelty is replacing a trained sign predictor with an unsupervised iterative selector. If the iteration frequently picks the wrong branch, the oracle cosine similarities (0.74–0.87) do not support the claimed mechanism.\n\nSecond, the NP-MSGLA update in Eqs. (10)–(12) does not actually implement the law-of-sines formula derived in Eq. (9); it uses the estimated noise phase in a spectral-subtraction operation. That can be a legitimate approximation, but the paper should say so instead of implying the update directly applies the derivation.\n\nThird, the empirical gains in Tables 2 and 3 are within a few hundredths of PESQ and SI-SNR points, and some comparisons are negative (e.g., NM-MSGLA's SI-SNR on WSJ0 vs. the sign predictor). No variance or significance tests are reported. The practical advantage over the sign predictor is not clearly established.\n\nMinor: a 'multi-mource' typo, and the 'first law of sines' claim should be softened.\n\nWho is this for? People working on phase estimation in speech enhancement or source separation. It is a modest but real step, and the paper is honest about the empirical nature of the convergence observation. I would send it to referees and ask for a proper convergence analysis (even a fixed-point or contraction argument for a simplified model, plus per-bin agreement with the two candidates), and for confidence intervals or repeated-seed results. The central claim is plausible but not yet demonstrated.","headline":"A genuine algorithmic combination of GLA-style consistency and geometric phase constraints, but the claimed convergence to the two candidates—the whole novelty—is an unsupported empirical observation and the measured gains are marginal.","tokens_in":9921,"tokens_out":3963,"would_cite":false,"duration_ms":47600,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that alternating speech-noise projections converge within about five iterations to one of the two valid phase candidates, resolving the sign ambiguity without a trained sign predictor.","keywords":["speech enhancement","phase estimation","Griffin-Lim algorithm","STFT consistency","geometric constraints","law of sines","sign ambiguity","additive noise"],"falsifier":"Run NM-MSGLA from several random phase initializations on one fixed noisy utterance and record where the five-iteration output lands: if it ever lands away from $P_Y + |\\Delta\\hat{P}|$ and $P_Y - |\\Delta\\hat{P}|$, if the chosen candidate changes with initialization, or if continuing to fifty iterations changes the answer, then the claimed unsupervised sign resolution cannot be relied on.","tokens_in":8916,"feed_emoji":"🎙️","tokens_out":21872,"duration_ms":180531,"temperature":0.7,"pith_summary":"The paper sets out to show that the sign ambiguity in geometry-based phase estimation can be resolved by iteration instead of by a trained classifier. Its multi-source Griffin-Lim algorithm (MSGLA) alternates STFT-consistency projections between the clean and noise spectrograms, and the authors report that after about five iterations the estimated speech phase settles on one of the two candidates that the law of cosines allows, or on one of two candidates from a new law-of-sines formulation that uses the noise phase. If this convergence is dependable, the DNN-based sign predictor used in earlier geometry-based systems becomes unnecessary. In oracle experiments and on the VB-DMD and WSJ0-CHiME3 benchmarks, the method matches or slightly beats direct phase estimation and sign prediction, with its clearest advantage in background-noise suppression.","feed_headline":"Five iterations replace the trained sign predictor in phase estimation","feed_subtitle":"Geometry plus STFT consistency settles which way the phase difference points, with no sign classifier to train.","key_machinery":"The load-bearing object is the MSGLA alternating-projection loop, a sequence of three steps per iteration: a GLA-style consistency projection of the speech spectrogram, a GLA-style projection of the noise residual, and a geometric re-derivation of the speech phase from $H_Y - \\hat{A}_Z e^{j\\tilde{P}_Z}$ (NM-MSGLA) or from the law-of-sines relation (NP-MSGLA). The law-of-sines relation is the paper's new geometric ingredient: it turns known speech magnitude and noise phase into two candidate clean phases per time-frequency bin, a formulation the authors say has not been used before. The work of the loop is to convert sign selection from a classification problem into a fixed-point problem: the paper's claim is that the consistency constraints alone select the correct candidate, so the entire geometry-based pipeline reduces to a magnitude estimator followed by a small number of cheap projections.","core_discovery":"The paper's central claim is that the sign choice in Equation (4) — whether to add or subtract the absolute phase difference $|\\Delta\\hat{P}|$ — need not be learned. By alternately projecting the speech magnitude $\\hat{A}_X$ and the noise residual $H_Y - \\hat{A}_X e^{j\\tilde{P}_X}$ through the STFT consistency operator $STFT \\circ iSTFT$, and re-deriving the speech phase from the additive model $H_Y = H_X + H_Z$, the estimate $\\hat{P}_X^{(n)}$ is observed to converge within a few iterations to either $P_Y + |\\Delta\\hat{P}|$ or $P_Y - |\\Delta\\hat{P}|$ (NM-MSGLA). The companion NP-MSGLA variant instead reconstructs phase from the law of sines: given the estimated speech magnitude and noise phase, the clean phase must be one of the two values $\\sin^{-1}(\\frac{A_Y}{\\hat{A}_X}\\sin(P_Y - \\hat{P}_Z)) + \\hat{P}_Z$ or $\\pi - \\sin^{-1}(\\frac{A_Y}{\\hat{A}_X}\\sin(P_Y - \\hat{P}_Z)) + \\hat{P}_Z$, and the same alternating-projection mechanism selects between them. The authors report that with oracle magnitudes or oracle noise phase, phase reconstruction closely tracks the ground truth, and in blind settings MSGLA matches or slightly outperforms direct phase estimation, plain GLA, and the DNN-based sign predictor, with the largest gains on background-suppression metrics (SI-SNR and CBAK).","pith_inferences":["The convergence claim is empirical, so a fixed-point analysis of the alternating projections is the natural next step: a proof that the two geometric candidates are the only stable attractors would convert the method from an observation into a theorem.","The additive-model recipe generalizes to more than two components, so the same alternating-projection idea could attack sign ambiguity in multi-talker mixtures, which has the same binary structure per time-frequency bin.","NP-MSGLA suggests a simpler training objective for single-channel enhancement: a model that outputs a speech magnitude mask plus a noise phase estimate could replace full complex-spectrogram prediction.","Because STFT consistency depends on the analysis window, the number of iterations needed and the reliability of sign selection may vary with window length and hop size; a sweep over those settings would test whether the 'around five iterations' behavior holds regardless of the window."],"forward_implications":["In geometry-based phase estimation, the trained sign predictor can be dropped: roughly five alternating projections select the sign, removing the misclassification errors that near-random sign targets cause.","Noise phase becomes a usable cue for the first time: NP-MSGLA derives phase candidates from the law of sines, and it is most reliable in low-energy regions where clean speech phase estimation fails.","Oracle experiments indicate that noise magnitude accuracy matters more than speech magnitude accuracy for phase reconstruction, a result the paper highlights as pointing magnitude estimation toward the noise component.","On VB-DMD and WSJ0-CHiME3, both MSGLA variants match or slightly outperform direct phase estimation and the DNN sign predictor, with NP-MSGLA posting the best SI-SNR and CBAK scores on VB-DMD."],"supporting_citations":[{"why":"Supplies the Griffin-Lim alternating-projection procedure that MSGLA adapts to the two-source additive model.","marker":"[33]"},{"why":"Introduces the law-of-cosines geometry that produces two phase candidates and the sign ambiguity MSGLA resolves.","marker":"[27]"},{"why":"The DNN-based sign predictor that MSGLA is designed to replace and that it matches or beats in background suppression.","marker":"[28]"},{"why":"Provides the law of sines and cosines identities from which the two candidate phase formulas are derived.","marker":"[31]"},{"why":"Defines the STFT consistency constraint that the GLA-style projections enforce at every iteration.","marker":"[34]"},{"why":"The direct phase estimation model whose learned error patterns motivate NP-MSGLA, since noise phase is easier to estimate in low-energy regions.","marker":"[20]"},{"why":"The TF-GridNet backbone used to produce the estimated speech magnitude, noise magnitude, and noise phase in the blind experiments.","marker":"[21]"}],"fun_headline_variants":["Geometry and consistency constraints replace sign prediction","Iterative phase estimation without a trained sign classifier","Alternating projections pick the correct phase sign","Multi-source Griffin-Lim eliminates the sign predictor","STFT consistency settles phase sign ambiguity in five steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the empirical observation that after about five iterations the alternating projections converge to one of the two geometrically valid phase candidates; if the loop can oscillate, land on a spurious point, or depend on initialization, the unsupervised sign resolution that replaces the trained classifier gives way.","fun_headline_variants_meta":{"raw":{"variants":["Geometry and consistency constraints replace sign prediction","Iterative phase estimation without a trained sign classifier","Alternating projections pick the correct phase sign","Multi-source Griffin-Lim eliminates the sign predictor","STFT consistency settles phase sign ambiguity in five steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000301,"raw_usage":{"total_tokens":1775,"prompt_tokens":1026,"completion_tokens":749,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":680}},"tokens_in":642,"tokens_out":749,"duration_ms":7928,"temperature":1.0,"reasoning_tokens":680,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:34:48.417055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run NM-MSGLA from several random phase initializations on one fixed noisy utterance and record where the five-iteration output lands: if it ever lands away from $P_Y + |\\Delta\\hat{P}|$ and $P_Y - |\\Delta\\hat{P}|$, if the chosen candidate changes with initialization, or if continuing to fifty iterations changes the answer, then the claimed unsupervised sign resolution cannot be relied on.","supporting_citations":[{"cited_title":"Signal estimation from modified short-time fourier transform,","cited_arxiv_id":null,"evidence_quote":"Supplies the Griffin-Lim alternating-projection procedure that MSGLA adapts to the two-source additive model."},{"cited_title":"Phase estimation for signal reconstruction in single-channel source separation,","cited_arxiv_id":null,"evidence_quote":"Introduces the law-of-cosines geometry that produces two phase candidates and the sign ambiguity MSGLA resolves."},{"cited_title":"Deep learning based phase reconstruction for speaker separation: A trigonometric perspective,","cited_arxiv_id":null,"evidence_quote":"The DNN-based sign predictor that MSGLA is designed to replace and that it matches or beats in background suppression."},{"cited_title":"Stewart, L","cited_arxiv_id":null,"evidence_quote":"Provides the law of sines and cosines identities from which the two candidate phase formulas are derived."},{"cited_title":"Explicit consistency constraints for stft spectrograms and their application to phase reconstruction","cited_arxiv_id":null,"evidence_quote":"Defines the STFT consistency constraint that the GLA-style projections enforce at every iteration."},{"cited_title":"MP-SENet: A speech enhancement model with parallel denoising of magnitude and phase spectra,","cited_arxiv_id":null,"evidence_quote":"The direct phase estimation model whose learned error patterns motivate NP-MSGLA, since noise phase is easier to estimate in low-energy regions."},{"cited_title":"Tf- gridnet: Making time-frequency domain models great again for monaural speaker separation,","cited_arxiv_id":null,"evidence_quote":"The TF-GridNet backbone used to produce the estimated speech magnitude, noise magnitude, and noise phase in the blind experiments."}],"review_version":1}