{"id":"ef6b3852-d274-4f0c-84f2-a37b244befc2","arxiv_id":"2501.05557","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An ADMM-based joint estimation algorithm that reconstructs full-band magnitude and phase from mel-spectrograms, achieving better SCM, PESQ, ESTOI, and spectral convergence with fewer iterations than prior iPALM and cascaded methods.","lead":"This paper proposes a faster and more accurate way to reconstruct an audio waveform from its mel-spectrogram, a compressed spectral representation used in speech synthesis, by solving the problem with the ADMM optimization algorithm. The method jointly estimates the missing full-band magnitude and phase in closed-form updates, beat existing signal-processing baselines on speech and foley sounds, and needs no training.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on unproven convergence of nonconvex multi-block ADMM; the paper reports no primal/dual residuals or stationarity checks, so the observed gains over iPALM could plausibly stem from early stopping or tuned hyperparameters rather than the ADMM formulation itself.","rationale":"I agree with the reader that the most load-bearing weakness is the absent convergence analysis for the nonconvex multi-block ADMM, and I would keep the conditional verdict. The mathematical derivation of the ADMM updates is sound: the X-update in (15) is the correct proximity operator for || |X|-Y ||^2, the W-update in (17) is the correct least-squares solve, and the Y-update in (20) follows from projecting the weighted average onto the nonnegative orthant. The complexity claim is plausible if both algorithms use a precomputed FxT matrix multiplication, though the W-update's dense inverse could be more expensive unless a Woodbury identity is used. The honest concern is that the paper presents empirical SCM curves and boxplots but no evidence that the iterates actually converge to a stationary point of the original problem. Because the paper describes itself as based on a rigorous optimization algorithm, the burden is on the authors to show that the ADMM iterates satisfy some stationarity condition. The proposed residual test would settle whether the improvement is a genuine convergence property or a byproduct of early stopping and hyperparameter selection. This does not change the reader's conditional verdict; it reinforces the conditions under which the claim should be accepted.","tokens_in":9524,"tokens_out":14291,"duration_ms":138134,"concrete_test":"Run Algorithm 2 on a fixed TIMIT utterance with the reported (lambda,rho)=(5000,0.1), and record the primal residuals ||Z-X|| and ||Y-W||, the dual residuals rho*||V_k-V_{k-1}|| and rho*||U_k-U_{k-1}||, and the norm of a proximal-gradient mapping for problem (7) at iterations 1, 10, 100, and 500. If these residuals do not decay to a small fraction of their initial values, the algorithm is not converging to a stationary point, and the SCM advantage may reflect early stopping; repeat with rho=0.02 and rho=0.5 to assess sensitivity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's central claim is that Algorithm 2 (ADMM-Joint) solves problem (7) and outperforms iPALM, with a 5x iteration-count advantage. This claim rests on the assumption that the nonconvex, nonsmooth ADMM iterates converge to a useful stationary point. The paper provides no convergence analysis, and the objective J(X,Y) has a gradient w.r.t. X that is not Lipschitz at |X|=0 (the paper sets the derivative to zero there, as noted in Section II-A), so standard nonconvex ADMM theorems do not directly apply. The reported best value rho=0.1 is small, a regime where ADMM can oscillate; the paper only shows SCM curves and does not report primal residuals ||Z-X|| and ||Y-W||, dual residuals, or stationarity violations. Without such evidence, the observed improvement could be an artifact of early stopping or of hyperparameters tuned on the validation set, rather than a genuine property of the ADMM scheme.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an ADMM-based joint estimation method for mel-spectrogram inversion. It reformulates the joint full-band magnitude and phase reconstruction problem (7) as an ADMM problem (8), derives closed-form proximity-operator updates for X, W, Z, and Y (Section III-B, Eqs. (12)–(20)), and presents Algorithm 2. Experiments on TIMIT speech and DCASE2023 foley sounds compare ADMM-Joint with cascaded PG-GLA, ADMM-GLA, and the iPALM-based joint method [27], reporting improved SCM, PESQ, ESTOI, and full-band spectral convergence, with the claim that ADMM-Joint with 100 iterations is comparable to iPALM-Joint with 500 iterations (Section IV-B).","tokens_in":9737,"tokens_out":3158,"duration_ms":33140,"significance":"If the empirical claims hold, this is a useful training-free, signal-processing-based mel-spectrogram inversion method that reduces the iteration count of the prior joint estimation method by a factor of five while improving objective quality on speech and foley sounds. The derivation of the ADMM updates is largely correct, and the conditional-independence splitting is a sensible way to keep each subproblem simple. The experiments cover multiple metrics and two domains, which is a strength. However, the central comparison is weakened by the lack of any convergence or stationarity analysis for the nonconvex multi-block ADMM scheme, and by the fact that the primary metric SCM is essentially the objective being minimized; the external metrics (PESQ/ESTOI and full-band SC) partially mitigate this but do not remove the need for a rigorous assessment of whether the algorithm actually solves (7).","major_comments":[{"comment":"The central claim that ADMM-Joint solves problem (7) and converges to a useful stationary point is not supported by any convergence analysis. The objective (7) is nonconvex, and the gradient of J(X,Y) with respect to X is not Lipschitz at |X|=0, as the paper itself notes in Section II-A when setting the derivative to zero. Standard nonconvex ADMM theorems cited in Section III-A do not directly apply to this multi-block, non-Lipschitz setting. The paper reports no primal residuals ||Z-X|| or ||Y-W||, no dual residuals, and no stationarity-violation measures over iterations. Without such evidence, the observed advantage over iPALM at 100 iterations could be an artifact of early stopping or of the particular choice of ρ rather than a genuine convergence property. Please add residual/stationarity plots or a convergence analysis for Algorithm 2.","section":"Section III-B, Algorithm 2"},{"comment":"The comparison with iPALM-Joint may not be fully fair with respect to hyperparameters. Figure 1 shows that the proposed method's performance varies substantially with ρ, and the pair (λ,ρ)=(5000,0.1) is selected on a 100-utterance validation set. The iPALM-Joint baseline uses λ=10 taken from the prior work [27] without re-tuning on the same validation set, and its inertial parameter α is not specified. To support the claim of superiority, both methods should be tuned on the same validation data, or the authors should show that ADMM-Joint remains competitive across a range of ρ and that the chosen ρ is not a single fragile point.","section":"Section IV-A, Fig. 1"},{"comment":"The SCM metric defined in Eq. (21) is exactly the mel-reconstruction error ∥E|STFT(x̂)|−M∥/∥M∥, which is the term I(Y) minimized in problem (7). Thus SCM gains for ADMM-Joint partly reflect better minimization of the same objective, making SCM a circular measure of quality. The paper does report external quality measures (PESQ and ESTOI on speech, full-band SC on foley) that support the claim, but the main convergence-speed comparison in Fig. 2 is made in terms of SCM. Please clearly acknowledge this circularity, and preferably report the convergence comparison also in terms of an external metric or a stationarity measure.","section":"Section IV-B, Eqs. (7), (21)"},{"comment":"For the speech experiments, the paper does not report any statistical significance test for the differences in SCM, PESQ, or ESTOI between ADMM-Joint and iPALM-Joint. The foley results include a paired t-test (Section IV-C), but the speech boxplots are interpreted qualitatively. Given the claim of consistent improvement, please add a significance test or confidence intervals for the speech metrics as well.","section":"Section IV-B, Fig. 2-3"}],"minor_comments":[{"comment":"The abbreviation 'iPLAM' appears in the text and in Figure legends; it should be 'iPALM' consistently.","section":"Section IV-B"},{"comment":"The phrase 'spectral convergence on mel-spectrogarm' contains a typo: 'mel-spectrogram'.","section":"Section IV-A"},{"comment":"The top-left panel shows SCM versus λ only, while the other two panels show PESQ and ESTOI versus ρ; for completeness, please also show how PESQ and ESTOI vary with λ, since the selected λ=5000 is justified only by SCM.","section":"Section IV-A, Fig. 1"},{"comment":"Please report the exact p-value or t-statistic for the paired t-test on foley sounds, not merely the statement that significance was confirmed.","section":"Section IV-C"},{"comment":"The initialization of variables for both algorithms is not specified. Please state how Z, Y, V, U are initialized for ADMM-Joint and how Z, Zold, Y are initialized for iPALM-Joint, since initialization can affect convergence in nonconvex problems.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid signal-processing contribution, and the derivation of the ADMM updates is correct and easy to check. The main weakness is the absence of any convergence or stationarity analysis for a nonconvex multi-block ADMM scheme whose objective is non-Lipschitz near zero; the authors should either provide a proof under suitable assumptions or report residuals/stationality checks that empirically support the claimed behavior. The SCM circularity and the hyperparameter-tuning asymmetry are also load-bearing issues that can be addressed within the manuscript's scope. I would not recommend rejection, but the revision needs to substantiate the central convergence and fairness claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate new algorithm. The authors reformulate mel-spectrogram inversion as a joint estimation of full-band magnitude and phase, split variables so every ADMM subproblem has a closed-form proximal update, and show on TIMIT speech and DCASE2023 foley that it beats the prior iPALM joint method with roughly five times fewer iterations. The math is sound: the X, W, Z, and Y updates are the correct proximity operators, and the conditional-independence decomposition into (12a)-(12d) is valid. Experiments are reasonably clean: separate hyperparameter search, PESQ and ESTOI on speech, spectral convergence on foley, and a paired t-test for the foley results. The paper also honestly notes it still trails neural vocoders.\n\nSoft spots, in proportion. The main gap is the missing convergence analysis. The problem is nonconvex, and the gradient of J with respect to X is not Lipschitz at |X|=0 (the authors set the derivative to zero there), so standard nonconvex ADMM theorems do not directly apply. The paper cites ADMM's success elsewhere but proves nothing for this specific scheme. More concretely, no primal residuals ||Z-X|| and ||Y-W||, no dual residuals, and no stationarity-violation plots are reported, so it is hard to tell whether the 100-iteration advantage is genuine ADMM behavior or an artifact of early stopping combined with tuned hyperparameters. The sensitivity to rho shown in Fig. 1 is real, and the 5x claim may be dataset-dependent. The foley gains appear marginal (a few dB) though statistically significant. Lack of released code is a reproducibility annoyance, not a flaw in the method.\n\nNone of this is fatal. For this subfield, an empirical paper with correct derivations and consistent gains is a genuine contribution. The SCM metric is partly circular because it is exactly the objective the algorithm minimizes, but PESQ, ESTOI, and full-band spectral convergence are external and they move in the right direction. I would want a revision that reports residuals, a brief stationarity check, and ideally releases code, but the central empirical claim holds up on the evidence.\n\nWho should read this: people working on signal-processing-based phase reconstruction and mel-spectrogram inversion, and anyone building baselines against neural vocoders. It deserves a serious referee; I would not desk-reject it.","headline":"Solid ADMM-based mel-spectrogram inversion with correct closed-form updates and real gains over iPALM; missing convergence analysis and residual reporting, but worth reviewing.","tokens_in":10247,"tokens_out":3031,"would_cite":true,"duration_ms":27433,"reading_group":"yes","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 proposes an ADMM-based algorithm for mel-spectrogram inversion that jointly estimates full-band STFT magnitude and phase, achieving better reconstruction quality and roughly five times fewer iterations than the previous…","keywords":["mel-spectrogram inversion","phase reconstruction","alternating direction method of multipliers","nonconvex optimization","proximal splitting","speech synthesis","foley sound synthesis","full-band magnitude estimation"],"falsifier":"Run Algorithm 2 with the paper's recommended $(\\lambda, \\rho) = (5000, 0.1)$ on a fixed TIMIT utterance and monitor the primal residuals $\\|Z - X\\|$ and $\\|Y - W\\|$ across iterations; if these residuals do not decrease while SCM saturates, the assumed ADMM convergence is violated. A broader test: on a held-out corpus with random initializations, if ADMM-Joint's median SCM after 500 iterations is not better than iPALM-Joint's, or if its 100-iteration SCM does not match iPALM-Joint's 500-iteration SCM, the central speed and quality claims fail.","tokens_in":9321,"feed_emoji":"🎧","tokens_out":6284,"duration_ms":55203,"temperature":0.7,"pith_summary":"Mel-spectrogram inversion commonly works in two stages: first predict a full-band STFT magnitude from the mel-spectrogram, then run a phase-reconstruction algorithm such as Griffin–Lim. This paper's central claim is that those two steps should be optimized jointly, and that an alternating direction method of multipliers (ADMM) formulation does the joint estimation better and with far fewer iterations than the existing iPALM-based joint method. The authors report that their ADMM-based algorithm, iterated only 100 times, matches the quality the previous joint method reaches after 500 iterations, and that it improves spectral convergence, PESQ, and ESTOI on speech and full-band spectral convergence on foley sounds. The payoff is a training-free inversion method that is applicable to any signal yet competitive with classical signal-processing approaches.","feed_headline":"ADMM inverts mel-spectrograms better and 5x faster","feed_subtitle":"Jointly estimates magnitude and phase, beating iPALM and cascaded baselines on speech and foley sounds.","key_machinery":"The central object is the ADMM reformulation of problem (7): the full-band magnitude is split into two variables, one tied to the mel-spectrogram fidelity and one tied to the nonnegative magnitude constraint, and the STFT coefficient is split from a consistency variable constrained to the image of STFT. The workhorse is conditional independence among the blocks, which reduces the joint updates to four separate proximity-operator evaluations. The X-update uses the closed-form prox for magnitude least squares, $\\operatorname{prox}_{J(\\cdot,Y)/\\rho}(\\Psi) = (Y + \\rho|\\Psi|)/(1+\\rho)\\odot \\Psi/|\\Psi|$; the W-update solves $(\\lambda E^T E + \\rho I)^{-1}(\\lambda E^T M + \\rho \\Phi)$, with the inverse precomputable; the Z-update is STFT–iSTFT projection; and the Y-update is a soft-thresholded weighted combination $(|X| + \\rho \\Upsilon)_+/(1+\\rho)$.","core_discovery":"The paper builds a mel-spectrogram inversion algorithm, ADMM-Joint, by rewriting the joint full-band magnitude and phase estimation problem as a constrained optimization with auxiliary variables and then applying ADMM. Each of the four block updates becomes a proximal operator calculation: the STFT coefficients are updated by a magnitude-constrained least-squares prox, the auxiliary full-band magnitude by a closed-form ridge-like solve whose matrix inverse can be precomputed, the consistent-STFT variable by projection onto the STFT image, and the magnitude variable by a projection of a weighted combination of magnitude and dual variables. On TIMIT speech and DCASE2023 foley sounds, ADMM-Joint outperforms the cascaded PG-GLA and ADMM-GLA baselines and the iPALM-based joint method [27], with the 100-iteration ADMM run comparable to the 500-iteration iPALM run.","pith_inferences":["The paper leaves implicit that the same ADMM scheme could be unfolded into a trainable network, similar to deep Griffin–Lim; the closed-form prox updates are differentiable, so a learned version is a natural next step.","The reported sensitivity to $\\rho$ hints that an adaptive penalty schedule, which the authors do not test, could further reduce iteration counts and remove the manual tuning burden.","A rigorous convergence analysis for the nonconvex ADMM scheme would strengthen the empirical claim; until then, the method's reliability on out-of-distribution signals is an open question."],"forward_implications":["If the central claim holds, training-free mel-spectrogram inversion can be made roughly five times faster at equal quality, which matters for real-time and embedded speech synthesis pipelines.","Because the method requires no training data, it can be applied immediately to any signal class, including foley and environmental sounds where neural vocoders are rarely available.","The ADMM structure suggests that the same block-splitting strategy could be reused for other audio inverse problems that combine a redundancy constraint with a lossy perceptual transform.","The comparisons to cascaded methods indicate that the joint update's error mitigation, rather than the phase-only refinement, is what drives the quality gain."],"supporting_citations":[{"why":"Griffin–Lim algorithm; the baseline iterative phase reconstruction that cascaded methods build on.","marker":"[6]"},{"why":"ADMM-based phase recovery that motivates applying ADMM to mel-spectrogram inversion.","marker":"[8]"},{"why":"The existing iPALM-based joint full-band magnitude and phase estimation method that the proposed method compares against and seeks to improve.","marker":"[27]"},{"why":"The standard ADMM reference supplying the alternating minimization and dual-update structure.","marker":"[28]"},{"why":"Proximal operator machinery used to derive closed-form updates of each block.","marker":"[30]"},{"why":"iPALM algorithm, the optimization framework behind the baseline joint method.","marker":"[33]"}],"fun_headline_variants":["ADMM speeds up mel-spectrogram inversion 5x","Joint ADMM beats iPALM in mel inversion","5x faster mel inversion via ADMM joint estimation","ADMM improves mel inversion speed and quality","Mel inversion: ADMM outperforms cascaded baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's advantage assumes that multi-block ADMM converges to a good stationary point of this nonconvex problem within a few hundred iterations for arbitrary signals and without special initialization; the paper does not provide a convergence proof.","fun_headline_variants_meta":{"raw":{"variants":["ADMM speeds up mel-spectrogram inversion 5x","Joint ADMM beats iPALM in mel inversion","5x faster mel inversion via ADMM joint estimation","ADMM improves mel inversion speed and quality","Mel inversion: ADMM outperforms cascaded baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1377,"prompt_tokens":882,"completion_tokens":495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":419}},"tokens_in":498,"tokens_out":495,"duration_ms":4725,"temperature":1.0,"reasoning_tokens":419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:12:34.044815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 2 with the paper's recommended $(\\lambda, \\rho) = (5000, 0.1)$ on a fixed TIMIT utterance and monitor the primal residuals $\\|Z - X\\|$ and $\\|Y - W\\|$ across iterations; if these residuals do not decrease while SCM saturates, the assumed ADMM convergence is violated. A broader test: on a held-out corpus with random initializations, if ADMM-Joint's median SCM after 500 iterations is not better than iPALM-Joint's, or if its 100-iteration SCM does not match iPALM-Joint's 500-iteration SCM, the central speed and quality claims fail.","supporting_citations":[{"cited_title":"Signal estimation from modified short-time Fourier transform,","cited_arxiv_id":null,"evidence_quote":"Griffin–Lim algorithm; the baseline iterative phase reconstruction that cascaded methods build on."},{"cited_title":"Griffin–Lim like phase recov- ery via alternating direction method of multipliers,","cited_arxiv_id":null,"evidence_quote":"ADMM-based phase recovery that motivates applying ADMM to mel-spectrogram inversion."},{"cited_title":"Signal reconstruction from mel- spectrogram based on bi-level consistency of full-band magnitude and phase,","cited_arxiv_id":null,"evidence_quote":"The existing iPALM-based joint full-band magnitude and phase estimation method that the proposed method compares against and seeks to improve."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The standard ADMM reference supplying the alternating minimization and dual-update structure."},{"cited_title":"Proximal algorithms,","cited_arxiv_id":null,"evidence_quote":"Proximal operator machinery used to derive closed-form updates of each block."},{"cited_title":"Inertial proximal alternating linearized mini- mization (iPALM) for nonconvex and nonsmooth problems,","cited_arxiv_id":null,"evidence_quote":"iPALM algorithm, the optimization framework behind the baseline joint method."}],"review_version":1}