{"id":"77370b65-5d89-4067-bc19-2e66822e71ef","arxiv_id":"2607.18629","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"CS-ETS applies Lyapunov and detrended-fluctuation-analysis losses inside a Samba encoder, but its headline audio gains are confounded by a DTW alignment step not applied to baselines.","lead":"CS-ETS is a smaller EMG-to-speech model that adds two chaos-derived loss terms and a post-vocoder alignment step, reporting large gains on audio-quality metrics. The paper is worth reading as a case study in how evaluation choices—aligning generated audio to the target before scoring—can drive claimed improvements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Post-hoc DTW alignment is applied only to the proposed model, not baseline [5]; Table 3 shows PVA alone accounts for nearly all STOI/LSD gains, so headline audio improvements are evaluation artifacts rather than model superiority.","rationale":"The reader's weakest_assumption pinpoints exactly the load-bearing flaw: PVA is applied asymmetrically, and the ablation in Table 3 shows PVA alone produces nearly all the STOI/LSD improvement. My stress-test confirms this is the decisive issue. The paper's strongest evidence outside the audio metrics is the WER ablation (41.26% vs 50.37% without chaotic losses), which is a legitimate model-level comparison not directly confounded by PVA. However, the headline claims in the abstract and Section 4.1 are explicitly framed as 2.1x/4.7x/1.25x audio-quality improvements over [5], and those are the claims that would attract acceptance. Since the baseline does not receive PVA, those numbers cannot be attributed to the proposed architecture or losses. The fix is straightforward: apply PVA to all models or to none, and report both. Until then, the central claim is unsupported. I find no reason to change the reader's REJECT verdict; the concern lands as identified.","tokens_in":10528,"tokens_out":2898,"duration_ms":28184,"concrete_test":"Re-run Tables 1 and 3 applying the identical PVA procedure from Section 2.4 to the retrained baseline [5] before computing LSD/STOI/SI-SDR, and compare against CS-ETS with the same PVA. If the baseline's aligned scores approach CS-ETS (e.g., STOI rises from 0.13 to >0.5), the claimed 2.1x/4.7x/1.25x gains are artifacts of unequal evaluation. Also compare both models without PVA; the current Table 3 P1 vs P0 differences (LSD 1.92 vs 2.25, STOI 0.15 vs 0.13) are far smaller and would provide a fair basis for the model claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that CS-ETS improves audio quality by 2.1x LSD, 4.7x STOI, and 1.25x SI-SDR over baseline [5] is not supported because the comparison is asymmetric. Section 2.4's Post-Vocoder Alignment (PVA) is applied only to CS-ETS; baseline [5] is scored on raw vocoder output. Table 3 isolates this confound: row P1 (CS-ETS, no PVA, no chaotic losses) has STOI 0.15 and LSD 1.92; row P2 (same model, adding PVA only) jumps to STOI 0.61 and LSD 1.10. These jumps are essentially the entire reported improvement over [5] (STOI 0.13→0.61, LSD 2.25→1.07). Thus the headline audio-quality gains are driven by the oracle-alignment post-processor, not by the modified Samba encoder or the LER/MSDFA losses. Because the baseline is denied the same post-processing, the comparison is not a controlled test of the proposed architecture. The WER ablation in Table 3 does provide independent evidence that LER+MSDFA improve intelligibility at fixed model size (50.37%→41.26% without PVA), but the paper's quantitative audio-quality claims are confounded and cannot be accepted as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CS-ETS, an EMG-to-speech synthesis model combining a Samba-based encoder with two chaos-inspired auxiliary losses (Lyapunov Exponent Regularization, LER, and Multi-Scale Detrended Fluctuation Analysis, MSDFA), plus a Post-Vocoder Alignment (PVA) procedure that DTW-aligns generated audio to the target before computing frame-level metrics. The authors claim a 40.79% parameter reduction over the baseline [5] along with large improvements in LSD (2.1x), STOI (4.7x), and SI-SDR (1.25x), and report WER improvements from the chaotic losses in ablations.","tokens_in":10927,"tokens_out":4984,"duration_ms":56528,"significance":"If the comparisons were controlled, a 40.79% parameter reduction with better or comparable quality would be a useful contribution to EMG-to-speech synthesis, and the WER ablation in Table 3 (50.37% without losses to 41.26% with LER+MSDFA at fixed model size) provides some evidence that the proposed losses help intelligibility. The paper is also transparent in reporting ablations and hyperparameters for the core model. However, the headline audio-quality gains are not supported by the experiments as presented: PVA is applied only to the proposed model, and Table 3 shows that PVA alone accounts for nearly the entire LSD/STOI improvement. The WER margin over the baseline is small (41.26% vs 42.20%) with no significance testing, and the chaos-theoretic premise is asserted rather than demonstrated on the actual signals. The central quantitative claims therefore need substantial rework before the paper can be accepted.","major_comments":[{"comment":"The headline comparison is confounded by asymmetric evaluation. In Table 3, the baseline [5] (row P0) is evaluated without PVA: STOI 0.13, LSD 2.25. Row P1 (CS-ETS without PVA and without chaotic losses) has STOI 0.15, LSD 1.92. Row P2 adds only PVA and jumps to STOI 0.61, LSD 1.10. Thus PVA alone produces nearly the entire reported improvement over [5] (STOI 0.15→0.61, LSD 1.92→1.10). Since baseline [5] is not subjected to the same PVA post-processing, Table 1's 2.1x/4.7x/1.25x claims are not a valid model comparison. The authors must either apply PVA to the baselines or report unaligned metrics for both systems.","section":"§4.1 and Table 3; §2.4"},{"comment":"The validity of PVA as an evaluation step is not established. DTW alignment of the synthesized waveform to the target before computing LSD and STOI can mechanically improve frame-aligned distortion/intelligibility metrics even if the underlying speech content is of low quality, because it removes temporal misalignment that these metrics penalize. Table 3 P1→P2 shows the effect is large (STOI 0.15→0.61). The paper should validate PVA by, for example, applying it to the baseline model and to a poor-quality or random system, and showing that it does not artificially inflate scores. Without such a control, the post-alignment metrics cannot be interpreted as evidence of model quality.","section":"§2.4, Table 3"},{"comment":"At matched conditions without PVA and without chaotic losses, the proposed compact encoder is worse than the baseline [5]: WER 50.37% (P1) vs 42.20% (P0). The only WER evidence for the full CS-ETS over the baseline is the final row, 41.26% vs 42.20%, a 0.94-point difference with no significance test, confidence interval, or multiple-seed variance. Given that WER is the one metric not affected by PVA, this marginal improvement is insufficient to support the broad claim of 'significantly improved performance' over [5] at 40.79% fewer parameters. Statistical testing or training-seed variance should be reported.","section":"§4.3, Table 3 (P0 vs P1)"},{"comment":"The conclusion states 'Experiments on Gaddy and Klein datasets [4] validate...' but Section 3.1 describes training on a single dataset introduced in [4], and the Limitations section says 'we do not use multiple datasets in a noisy setting in this paper.' This is an internal contradiction, and the claim of validation on multiple datasets is unsupported. The conclusion should be corrected to refer to the single dataset actually used.","section":"§6 Conclusion; §3.1; §6 Limitations"},{"comment":"The paper repeatedly asserts that EMG and speech production exhibit 'deterministic chaos' and that the proposed losses 'capture chaotic dynamics,' but no quantitative analysis of the actual EMG/mel features is provided. No Lyapunov spectrum estimate on real signals, no surrogate data test, and no comparison of chaotic statistics before/after training. Without this, 'chaos-inspired' is a plausibility argument rather than a demonstrated property, and the losses are better described as nonlinear feature-matching regularizers. A concrete falsifiable test would be to compute the largest Lyapunov exponent or DFA scaling exponent on the actual EMG features and compare against surrogate time series.","section":"§2, §2.3.1, §2.3.2"}],"minor_comments":[{"comment":"Several hyperparameters are not given values: the Theiler window size, kmax, and the MSDFA scale set s. Only m=10 and tau=1 are specified. These affect the loss computation and should be reported for reproducibility.","section":"§2.3.1, Alg. 1; §2.3.2, Alg. 2"},{"comment":"Row labels such as 'CS-ETS w/o PV A+LER+MSDFA' are ambiguous: it is unclear whether this means 'without PVA, LER, and MSDFA' or 'without PVA but with LER+MSDFA.' Please write 'w/o PVA, LER, MSDFA' and similarly for other rows.","section":"§4.3, Table 3"},{"comment":"Typographical and notation inconsistencies: 'PV A' vs 'PVA', 'SISDR' vs 'SI-SDR', 'FLOPS' vs 'FLOPs'.","section":"Throughout"},{"comment":"The subjective MOS comparison (4.21 vs 3.98) is based on 10 raters with no confidence intervals or significance test. As reported, the difference may be within listener noise; please provide per-item variance or a paired test.","section":"§5 Subjective Analysis"},{"comment":"Reference [18] appears to duplicate reference [16] ('Nonlinear framework for speech bandwidth extension'). Please check and consolidate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a promising core idea — adding nonlinear regularity losses to a compact Samba-based ETS encoder — and the WER ablation for LER+MSDFA is suggestive. However, the central evaluation is currently not trustworthy because of the asymmetric PVA application, and the paper contains internal inconsistencies (e.g., the 'multiple datasets' conclusion vs the limitations). These issues are fixable with additional experiments and careful rewriting, so I recommend major revision rather than outright rejection. If the authors cannot show that PVA does not inflate scores, or if applying PVA to the baseline removes the reported gains, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a genuine idea buried under a misleading headline. The WER ablation is the credible part: removing both chaotic losses worsens WER from 41.26% to 50.37% at the same model size, and each loss individually contributes. That is independent evidence that LER/MSDFA provide useful supervision, and the 40.79% parameter reduction versus the retrained baseline is also real. If the paper had led with those results, it would have a solid, if modest, contribution to ETS.\n\nThe problem is the audio metrics. The 2.1x LSD, 4.7x STOI, and 1.25x SI-SDR gains over baseline [5] are not a fair comparison. Section 2.4 applies PVA only to CS-ETS; baseline [5] is scored on raw vocoder output. Table 3 makes the confound explicit: adding PVA alone to CS-ETS without any chaotic losses jumps STOI from 0.15 to 0.61 and LSD from 1.92 to 1.10 — nearly the entire reported improvement over [5]. That is an evaluation artifact, not model superiority. The paper is transparent enough to show this, which I appreciate, but the abstract and Section 4.1 still present the unfair numbers as headline claims.\n\nOther soft spots, in proportion: no error bars, single-speaker dataset, hyperparameters selected on the same evaluation set, and no code or data released. The claim that EMG is chaotic is asserted rather than demonstrated, though the losses could be viewed as heuristic regularizers regardless. The subjective MOS (10 listeners, no significance test) is weak but not central.\n\nWho is this for? Someone working on silent speech interfaces or ETS specifically. The WER result and the compact encoder are worth knowing, and PVA could become a useful evaluation protocol if applied fairly to all systems. But as it stands, the main comparison is not supportable, and I would not cite the audio-quality numbers.\n\nRecommendation: send it to peer review, but the authors need to re-run baselines with PVA, report variance, and either release code or provide aligned evaluation scripts. The core idea deserves referee time; the current quantitative claims do not.","headline":"The WER results look real, but the headline audio-quality claims collapse once you see that DTW alignment is applied only to the proposed model.","tokens_in":11406,"tokens_out":1639,"would_cite":false,"duration_ms":21456,"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":"Chaos-based losses plus a Mamba-style encoder can cut EMG-to-speech model size by 40.79% while improving reconstructed speech quality.","keywords":["EMG-to-speech synthesis","silent speech interface","Lyapunov exponent regularization","detrended fluctuation analysis","chaos-inspired loss","Mamba state-space model","post-vocoder alignment","speech quality metrics"],"falsifier":"Score the retrained baseline with the same PVA pipeline used for CS-ETS. If its STOI rises close to 0.61 and LSD falls near 1.07, the claimed superiority of the chaos-inspired architecture evaporates; conversely, if the gap persists, the claim survives. Also run the LER/MSDFA losses on shuffled or phase-randomized targets to check they measure temporal structure rather than global energy.","tokens_in":10410,"feed_emoji":"🎙️","tokens_out":3862,"duration_ms":77021,"temperature":0.7,"pith_summary":"CS-ETS is an EMG-to-speech (ETS) system built on a Samba-style state-space encoder with sliding-window attention, trained with two new losses derived from chaos theory: Lyapunov Exponent Regularization (LER), which matches local divergence rates of nearby trajectories, and Multi-Scale Detrended Fluctuation Analysis (MSDFA), which matches fractal long-range correlations. The paper claims that this combination yields a 40.79% smaller model than the strongest prior baseline while improving spectral reconstruction (LSD 1.07 vs 2.25), intelligibility (STOI 0.61 vs 0.13), and noise robustness (SI-SDR -33.41 vs -41.96). It also introduces Post-Vocoder Alignment, a DTW-based step that time-aligns vocoder output to the target so that frame-level metrics can be computed at all. A sympathetic reader would care because it suggests silent-speech interfaces can be both more compact and more accurate, and that chaotic temporal structure is a useful training signal rather than noise.","feed_headline":"Chaos losses shrink EMG-to-speech model by 41%","feed_subtitle":"Lyapunov-exponent and fractal-scaling losses let a smaller encoder match a larger transformer on silent speech.","key_machinery":"The load-bearing machinery is the pairing of a cheap recurrent-style encoder with two loss terms that compare predicted and target mel-spectrograms through nonlinear-dynamics summaries: the largest Lyapunov exponent (estimated by delay embedding, nearest-neighbor tracking, and log-divergence slope) and the DFA scaling exponent per frequency channel. Because these summaries are invariant to frame-by-frame spectral error, they give the model a training signal about temporal instability and long-range correlation. Post-Vocoder Alignment (DTW on cosine distance between STFT frames, with inverse STFT reconstruction) is what lets the authors compute LSD/STOI/PESQ at all and therefore produces the","core_discovery":"The paper's central claim is that EMG-to-speech synthesis benefits from explicitly supervising chaotic dynamics. On one speaker's 19-hour facial-EMG dataset, CS-ETS — a four-layer Mamba+SWA+MLP encoder with a convolutional front-end — trained with LER and MSDFA losses surpasses the prior best retrained baseline on WER (41.26% vs 42.20%) and on acoustic metrics enabled by Post-Vocoder Alignment, using 32.03M vs 54.10M parameters. The ablation table attributes most of the frame-level metric jump to the alignment procedure (STOI 0.15 to 0.61), and the chaotic losses to a smaller WER improvement (from 50.37% to 41.26% when both are added), so the paper's own numbers show chaos supervision chiefl","pith_inferences":["The headline quality ratios likely overstate model improvement: Table 3 shows STOI jumping from 0.15 to 0.61 when PVA is added without the chaos losses, so part of the 4.7x gain is an evaluation artifact. A direct test would re-score the baseline with PVA.","LER/MSDFA's marginal benefit seems concentrated in WER (50.37% to 41.26%) with only slight LSD movement; this pattern is consistent with these terms acting as a regularizer that stabilizes temporal dynamics rather than proof that EMG is deterministic chaos.","A natural extension is to apply the same losses to streaming or multi-speaker ETS, where the paper's single-speaker, noise-free setting leaves open whether the chaotic supervision generalizes."],"forward_implications":["ETS models can be compressed by about 41% (encoder from 44.03M to 21.96M parameters) and still match or beat a transformer baseline, making silent-speech interfaces more feasible on-device.","Frame-level metrics become available for ETS once Post-Vocoder Alignment is applied, so future work can report STOI/LSD/PESQ rather than relying on WER alone.","Chaos-derived losses may transfer to other speech or bio-signal mapping tasks where jitter and aperiodicity matter, since LER and MSDFA are not tied to EMG specifically.","The 13.33% FLOPs reduction (2.40G to 2.08G) supports real-time use at the same RTF, which is useful for wearable silent-speech systems."],"fun_headline_variants":["Chaos losses shrink EMG-to-speech model by 41% and improve WER","First chaos-supervised EMG-to-speech model uses 41% fewer parameters","Chaos losses drop EMG-to-speech WER from 50% to 41%","41% smaller EMG-to-speech model with chaos losses outperforms baseline","Chaos-supervised EMG-to-speech: 41% fewer parameters, better WER"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"That Post-Vocoder Alignment gives valid frame-level scores — if DTW matching the generated waveform to the target inflates LSD/STOI, the headline 2.1x/4.7x/1.25x improvements over a baseline scored without alignment are artifacts rather than model gains.","fun_headline_variants_meta":{"raw":{"variants":["Chaos losses shrink EMG-to-speech model by 41% and improve WER","First chaos-supervised EMG-to-speech model uses 41% fewer parameters","Chaos losses drop EMG-to-speech WER from 50% to 41%","41% smaller EMG-to-speech model with chaos losses outperforms baseline","Chaos-supervised EMG-to-speech: 41% fewer parameters, better WER"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000334,"raw_usage":{"total_tokens":1706,"prompt_tokens":774,"completion_tokens":932,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":819}},"tokens_in":518,"tokens_out":932,"duration_ms":13181,"temperature":1.0,"reasoning_tokens":819,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T14:47:33.921437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Score the retrained baseline with the same PVA pipeline used for CS-ETS. If its STOI rises close to 0.61 and LSD falls near 1.07, the claimed superiority of the chaos-inspired architecture evaporates; conversely, if the gap persists, the claim survives. Also run the LER/MSDFA losses on shuffled or phase-randomized targets to check they measure temporal structure rather than global energy.","supporting_citations":[],"review_version":1}