{"id":"1a235050-10c2-4df0-b424-eef40361d817","arxiv_id":"2602.22417","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ADDSE performs speech enhancement by absorbing discrete diffusion over neural audio codec tokens, reaching competitive non-intrusive quality and low-SNR robustness in few sampling steps.","lead":"A new system, ADDSE, cleans speech by using a diffusion model to fill in masked 'codes' from a neural audio codec, instead of processing raw waveforms. It reports good non-intrusive quality scores at low signal-to-noise ratios with very few sampling steps, but poor standard signal-reconstruction metrics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported competitiveness rests on test-set-selected Nsteps and non-intrusive predictors; a fixed operating point is needed before the claim is reliable.","rationale":"The reader's weakest assumption combines two related points: non-intrusive metrics may not be trustworthy arbiters, and test-set selection of Nsteps may inflate ADDSE's scores. The second is the most load-bearing because it is a concrete protocol flaw that can be fixed and directly affects the 'competitive' and 'few sampling steps' claims. The first is also important but more debatable; the paper explicitly limits its claim to non-intrusive metrics, so it is not internally wrong, only scientifically weak. I did not find a deeper internal inconsistency in the diffusion formulation: RQDiT is a plausible conditioned absorbing-diffusion model, and the training objective follows [39]. Independent support includes cross-corpus test sets and a sweep over model sizes. The conditional verdict is appropriate; a stricter rejection would be too strong because the method may still be competitive under a fair protocol, and validation of that protocol is cheap. Hence no change to the reader's CONDITIONAL verdict.","tokens_in":11757,"tokens_out":4646,"duration_ms":45580,"concrete_test":"Re-run the comparison with one fixed operating point per model: select Nsteps on a held-out validation set (or fix Nsteps=16 for all ADDSE systems), then report mean and 95% bootstrap confidence intervals over the 1000 test mixtures for every metric at that Nsteps, and also report the Nsteps that maximizes each metric to expose disagreement. If ADDSE no longer beats Conv-TasNet/SGMSE+ on non-intrusive metrics at the fixed Nsteps, or if PESQ/SDR remain far below baselines, the 'competitive' claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is explicitly scoped to non-intrusive objective metrics, so the decisive question is whether the reported comparison is a fair estimate of ADDSE quality. That condition is not established. Table 1's caption states 'For the ADDSE systems, the best Nsteps is selected' on the test set, and Figure 3 shows that the best Nsteps is metric-dependent: NISQA peaks at 8, the other non-intrusive metrics mostly plateau or improve through 16, and some continue to 1024. If different cells use different test-set-chosen Nsteps, the table does not describe a single runnable system, and the 'few sampling steps' conclusion is not actually what was measured. In addition, all ADDSE entries are single-run averages without confidence intervals, and the non-intrusive predictors are used as arbiters despite the method's much lower PESQ/ESTOI/SDR. The fact that ADDSE-XL exceeds the Clean-NAC upper bound on DNSMOS/NISQA (Libri-TUT) underscores that these predictors can reward codec-like output independently of enhancement fidelity. This is not an internal inconsistency, but it leaves the headline empirical claim underdetermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADDSE, a speech enhancement method that models the conditional distribution of clean speech codec tokens given noisy speech codec tokens using absorbing discrete diffusion in the latent space of a neural audio codec (NAC). The authors introduce RQDiT, a non-autoregressive hierarchical architecture that processes RVQ codebooks along frame and depth dimensions, and train it with a denoising cross-entropy objective. Inference starts from a fully masked state and unmasks clean codes with a few sampling steps. The method is evaluated on two cross-corpus test sets (Libri-TUT and Clarity-FSD50K) against waveform-domain, STFT-domain, and latent-domain baselines. The paper reports that ADDSE achieves competitive non-intrusive metrics, especially at low SNRs and with few sampling steps, while its intrusive metrics (PESQ, ESTOI, SDR) are substantially lower than those of waveform-domain baselines.","tokens_in":12119,"tokens_out":3190,"duration_ms":32478,"significance":"If the empirical claims hold, the paper makes a useful contribution by adapting absorbing discrete diffusion to speech enhancement, a formulation that has not been applied to this task before. The RQDiT architecture is a sensible adaption of RQ-Transformer and diffusion Transformers to hierarchical RVQ codes, and the use of the NAC codebook entries as token embeddings is elegant. The authors provide reproducible code and audio examples, and the training objective is well grounded in the published discrete-diffusion literature. The central novelty is the combination of a NAC latent space with a time-independent absorbing discrete diffusion model, with the potential for efficient low-step inference. However, the evidence supporting the headline 'competitive performance' claim is weakened by methodological choices in the evaluation, specifically the test-set-based selection of the number of sampling steps and the reliance on non-intrusive metrics as the primary arbiters.","major_comments":[{"comment":"The caption states 'For the ADDSE systems, the best Nsteps is selected.' Selecting Nsteps on the test set for each row means the reported numbers do not correspond to a single, fixed operating point of any ADDSE system. Figure 3 shows that the optimal Nsteps is metric-dependent: NISQA peaks at Nsteps=8, while DNSMOS, UTMOS, and SCOREQ generally improve or plateau through 16 and, for some systems, keep improving at 1024. Thus the DNSMOS and NISQA entries in Table 1 may come from different Nsteps values, so the table does not describe a runnable system and the 'few sampling steps' conclusion is not what was actually measured. The comparison should either fix Nsteps for all systems (or at least per dataset) based on a validation split, report the full Nsteps curves as primary evidence, or explicitly disclose the selected Nsteps and justify why test-set selection does not inflate the reporte","section":"5, Table 1 caption"},{"comment":"All results are single-run averages with no confidence intervals, significance tests, or variability estimates. Statements such as 'all ADDSE systems, including the small 4 M-parameter ADDSE-XS, outperform Conv-TasNet and SGMSE+ on DNSMOS and NISQA' cannot be assessed for reliability without a measure of uncertainty. Given that the differences between some systems are small (e.g., DNSMOS differences of 0.02–0.05), the ranking could easily change across runs or test subsets. The authors should report confidence intervals (e.g., bootstrap over utterances) and, where claims of superiority are made, paired significance tests.","section":"5, Table 1"},{"comment":"The central claim of competitive performance rests on non-intrusive metrics (DNSMOS, NISQA, UTMOS, SCOREQ), while the paper's own intrusive metrics (PESQ, ESTOI, SDR) are consistently much poorer for ADDSE than for waveform-domain baselines. The paper explains this by the NAC not reconstructing phase, but this does not by itself establish that the non-intrusive metrics are appropriate arbiters. In fact, ADDSE-XL exceeds the Clean-NAC upper bound on DNSMOS and NISQA on Libri-TUT (DNSMOS 3.76 vs 3.75; NISQA 3.43 vs 3.33), which suggests that these predictors can reward codec-like output regardless of enhancement fidelity. The manuscript should either provide evidence that the non-intrusive differences correspond to perceptually meaningful improvements (e.g., a listening test, or at least a correlation analysis with a reference-aware metric), or reframe the contribution as a codec-based gen","section":"5, Table 1; Figure 3; Section 5"}],"minor_comments":[{"comment":"The notation '1/λ δ_{cλ M}' is ambiguous. The indicator should be written as a function of the observed masked code, e.g., 1_{cλ,i = M}, and the expected value should be over the forward process transition. Please clarify.","section":"Eq. (7)"},{"comment":"The variable t is used both for the continuous time in Eq. (4) and for the scaled time in Eq. (9). This can confuse readers. Use a distinct symbol, e.g., τ, for the rescaled time.","section":"Sec. 2.2, Eq. (9)"},{"comment":"The description of RQDiT is clear overall, but it would help to state explicitly that the depth-DiT processes all depths for a given frame in parallel, and to specify the number of layers and heads for each DiT in the architecture figure or caption (the values 12 layers and 12 heads are only given in Sec. 4.3).","section":"Sec. 3.2"},{"comment":"The caption says 'best NISQA is achieved at Nsteps = 8, while the remaining metrics plateau at Nsteps = 16', but the curves for DNSMOS and SCOREQ continue to rise beyond 16 for several ADDSE sizes. The text understates this, and the claim of 'few sampling steps' should be qualified accordingly.","section":"Sec. 5, Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The main technical contribution—applying absorbing discrete diffusion with a hierarchical non-autoregressive transformer to codec tokens—is sound and worth publishing if the evaluation is brought up to standard. The decisive issue is that the reported 'competitive' performance is based on test-set-selected Nsteps and non-intrusive predictors, and the paper's own intrusive metrics tell a much less favorable story. I would like the editor to weigh whether the authors' planned revision addresses the operating-point problem and provides a more balanced metric presentation. If the authors cannot provide confidence intervals or a validation-based selection of Nsteps, the headline claim should be weakened accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: this is a competent, clearly-written application of absorbing discrete diffusion to codec-token speech enhancement, with a new two-axis DiT architecture, but the headline competitive claim rests on test-set-selected sampling steps and non-intrusive metrics. It deserves a serious referee, not a desk reject.\n\nThe genuinely new piece is RQDiT: two DiTs over frame and depth dimensions, reusing the NAC codebook embeddings instead of learned embeddings. That is a sensible adaptation of RQ-Transformer to non-autoregressive masked prediction, and it is well explained. The paper also correctly imports the time-independent parameterization from [39], which lets the network reuse predictions when no new token was unabsorbed, and it shows a real NFE reduction (average 545 at Nsteps=1024). The training objective is coherent, the baseline set is decent (Conv-TasNet, BSRNN, SGMSE+, plus two NAC-latent baselines), and the two test sets with cross-corpus noise/speech generalization are a plus.\n\nThe soft spots are real. Table 1 selects the best Nsteps per ADDSE system on the test set; that makes the table describe no single runnable system and weakens the \"competitive\" read. Figure 3 shows the best Nsteps is metric-dependent — NISQA peaks at 8, others plateau at 16 or keep rising. So the \"few sampling steps\" statement is not exactly what was measured. There are also no confidence intervals or significance tests, and all ADDSE entries are single-run averages. The bigger conceptual issue is that the competitive claim is carried by non-intrusive predictors (DNSMOS, NISQA, UTMOS, SCOREQ) while intrusive metrics (PESQ, ESTOI, SDR) are far below waveform-domain baselines. When ADDSE-XL beats the Clean-NAC upper bound on DNSMOS/NISQA, the predictors are rewarding codec-like output regardless of enhancement fidelity. That doesn't make the method useless, but it means the paper's own evidence is underdetermined: the improvement could be partly an artifact of the predictors rather than better restoration.\n\nThat said, I don't think these problems are fatal. The framework is plausible, the architecture is a genuine contribution, and the paper is honest about its scope — it claims non-intrusive competitiveness, and it does show that. The fix is a fixed Nsteps policy (e.g., chosen by validation) and error bars, plus a more careful discussion of why non-intrusive metrics diverge from PESQ/SDR. I'd send this to peer review. The evaluation protocol needs work, but the core idea and architecture deserve referee time.","headline":"Competent absorbing-discrete-diffusion SE paper with a useful RQDiT architecture, but the competitive claim depends on test-set-selected steps and non-intrusive metrics; worth reviewing with a demand for a fixed protocol.","tokens_in":12533,"tokens_out":1855,"would_cite":true,"duration_ms":17509,"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":"Speech enhancement can be recast as generating clean audio-code tokens from noisy ones with absorbing discrete diffusion.","keywords":["speech enhancement","absorbing discrete diffusion","neural audio codec","residual vector quantization","non-autoregressive generation","diffusion transformer","low-SNR speech","masked diffusion"],"falsifier":"A MUSHRA listening test comparing ADDSE outputs against the noisy input and the codec-reconstructed clean signal would settle the central claim; if listeners do not prefer ADDSE over the noisy input, or if the reported non-intrusive gains disappear when Nsteps is chosen on a validation set rather than the test set, the competitive claim is falsified.","tokens_in":11677,"feed_emoji":"🎧","tokens_out":6734,"duration_ms":59894,"temperature":0.7,"pith_summary":"ADDSE treats speech enhancement as a conditional generation problem in the discrete token space of a neural audio codec: given the codec codes of a noisy waveform, it models the distribution of the clean waveform's codes using absorbing discrete diffusion, in which tokens are progressively unmasked from a fully masked state. The authors propose a non-autoregressive Transformer architecture, RQDiT, that processes the hierarchical multi-codebook structure of residual vector quantization along both time and codebook-depth dimensions. They report that, measured by non-intrusive speech-quality metrics (DNSMOS, NISQA, UTMOS, SCOREQ), the approach is competitive with state-of-the-art waveform-domain systems on two held-out datasets, particularly at low signal-to-noise ratios and with as few as 8–16 sampling steps. A practical payoff is that network predictions can be reused across sampling steps, cutting the effective number of function evaluations by roughly half at 1024 steps. If the approach holds, convolutional and STFT-domain diffusion for enhancement can be replaced by cheap, parallel token-level generation that leverages advances in neural codecs and discrete diffusion.","feed_headline":"Discrete diffusion clears noisy speech in as few as 8 steps","feed_subtitle":"Modeling clean-code distribution given noisy codes makes small token-level systems beat larger baselines at low SNR.","key_machinery":"Two ingredients carry the argument. First, the absorbing state M — a mask token appended to each codebook — defines the forward process that gradually overwrites clean codes with M; the reverse process then unmask tokens by sampling from the learned conditional distribution, under a log-linear noise schedule that makes the transition probabilities simple and allows the model's predictions to be reused when no token was unmasked in the previous step. Second, RQDiT, a non-autoregressive architecture with a frame-DiT and a depth-DiT: the frame-DiT processes the sum of codebook embeddings across depths, the depth-DiT refines each frame independently along the codebook-depth dimension, and adapti","core_discovery":"On the paper's own terms, the central discovery is that speech enhancement can be posed as a conditional discrete-diffusion problem over neural-audio-codec tokens: the model learns q_theta(c | c_lambda, c_tilde), the distribution of clean RVQ codes given partially masked clean codes and the noisy codes, optimized with the denoising cross-entropy (Eq. 10). The authors show that, with the non-autoregressive RQDiT architecture and a time-independent reparametrization of the score, sampling from the fully absorbed state produces enhanced speech that scores competitively with established waveform-domain baselines on non-intrusive metrics on two cross-corpus test sets, and that the advantage is la","pith_inferences":["Editorial inference: the poor intrusive-metric scores (PESQ, ESTOI, SDR) likely reflect codec phase and reconstruction limitations rather than a failure of the diffusion sampler; a direct comparison of ADDSE against codec-reconstructed clean speech on those metrics would isolate the codec's contribution to the gap.","Editorial inference: the same masked-diffusion machinery should transfer to other token-sequence restoration tasks — bandwidth extension, declipping, or packet-loss concealment — since the conditional distribution is over generic RVQ tokens.","Editorial inference: selecting the best Nsteps on the test set may inflate the reported numbers; re-running the comparison with Nsteps chosen on a held-out validation set would test whether the favorable ranking persists, especially against the strongest waveform-domain baselines.","Editorial inference: the relationship between DCE and model size suggests DCE on a small validation set could serve as a cheap, sampling-free proxy for selecting model hyperparameters before spending compute on full reverse sampling."],"forward_implications":["Token-level non-autoregressive diffusion is a viable alternative to STFT-domain diffusion for enhancement, with near-peak quality at 8–16 sampling steps.","Because the codec is frozen and shared, the enhanced output is already a 2 kbps code sequence, so enhancement and compression are obtained in one pass.","The prediction-reuse trick cuts the number of network evaluations roughly in half at 1024 nominal steps, making high-step-count sampling cheaper than time-dependent models.","A 4 M-parameter ADDSE-XS already beats two established baseline systems on DNSMOS and NISQA on both test sets, indicating the framework is not compute-hungry at the low end.","Scaling model size from 4M to 580M parameters monotonically lowers the denoising cross-entropy and improves non-intrusive metrics, a clean scaling signal for further capacity increases."],"fun_headline_variants":["Discrete diffusion over audio codes cleans noisy speech","Absorbing diffusion enhances speech in a few sampling steps","Speech enhancement with discrete diffusion and RQDiT","Modeling clean speech codes: diffusion on neural codecs","Fast non-autoregressive diffusion for low-SNR speech"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The competitive claim rests on trusting non-intrusive metrics (DNSMOS, NISQA, UTMOS, SCOREQ) as faithful measures of enhancement quality, because the paper's intrusive metrics are far worse for ADDSE than for waveform-domain baselines, and on the assumption that selecting each system's best number of sampling steps on the test set does not inflate the comparison.","fun_headline_variants_meta":{"raw":{"variants":["Discrete diffusion over audio codes cleans noisy speech","Absorbing diffusion enhances speech in a few sampling steps","Speech enhancement with discrete diffusion and RQDiT","Modeling clean speech codes: diffusion on neural codecs","Fast non-autoregressive diffusion for low-SNR speech"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1183,"prompt_tokens":648,"completion_tokens":535,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":392,"completion_tokens_details":{"reasoning_tokens":457}},"tokens_in":392,"tokens_out":535,"duration_ms":5546,"temperature":1.0,"reasoning_tokens":457,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:42:07.613989+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A MUSHRA listening test comparing ADDSE outputs against the noisy input and the codec-reconstructed clean signal would settle the central claim; if listeners do not prefer ADDSE over the noisy input, or if the reported non-intrusive gains disappear when Nsteps is chosen on a validation set rather than the test set, the competitive claim is falsified.","supporting_citations":[],"review_version":1}