{"id":"ef802b9a-0a84-49e7-8bc8-4c6ae1c8e5b1","arxiv_id":"2502.02019","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 48 kHz neural codec that models complex spectra without downsampling shows strong robustness to out-of-domain expressive speech at 24 kbps.","lead":"ComplexDec is a new neural audio codec that compresses speech in the complex frequency domain, avoiding the information loss of standard waveform codecs. In tests, it keeps high speech quality when processing expressive speech it was never trained on, which typical codecs struggle with despite much larger training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Compression ratio is never varied independently of representation and architecture, so the central causal mechanism ('information loss degrades out-of-domain robustness') is not established by the reported comparisons.","rationale":"I read the paper in good faith. The empirical phenomenon is convincingly demonstrated: ComplexDec maintains MOS around 3.9-4.2 out-of-domain while AudioDec/ScoreDec collapse to 1.5-2.5, with non-overlapping 95% CIs, and the spectrogram analysis in Fig. 2 shows harmonic preservation that correlates with perceptual quality. The same-bitrate comparison against AudioDec/ScoreDec is fair, and the inclusion of open-source Encodec/DAC is a useful reference. My concern is not with the existence of the robustness effect but with the paper's explanatory mechanism. The phrase 'information loss from codec compression degrades out-of-domain robustness' is presented as an argument and then supported only by comparing systems that differ simultaneously in representation, architecture, quantization, losses, training data, and sample rate. This is exactly the confound the reader identified. A controlled ablation that varies compression ratio while holding representation and architecture fixed would settle whether the mechanism is causal, and would strengthen the paper's central claim. Until then, CONDITIONAL is the right verdict, and no code release or error bars change that assessment. I therefore recommend no change to the reader's verdict.","tokens_in":8540,"tokens_out":3635,"duration_ms":36034,"concrete_test":"Settle the confound with a crossed ablation at constant 24 kbps: (A) AudioDec as published (waveform, 5x compression), (B) AudioDec with code dimension raised to 256 and the projector removed so the compression ratio is 1.25x (waveform, low compression), (C) ComplexDec as published (spectral, 1.25x), and (D) ComplexDec with encoder/decoder temporal downsampling by 4 so the compression ratio is 5x (spectral, high compression). Train all four on the VCTK-derived Valentini-clean set only and measure out-of-domain MOS/SI-SDR on the EARS test set. If (B) approaches (C) and (D) degrades relative to (C), the information-loss mechanism is supported. If (B) stays poor or (D) stays robust, the robustness is due to the spectral representation or the architecture, not the compression ratio.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is causal: information loss from temporal and dimensional compression degrades out-of-domain robustness (Sec. I, II.B, IV). The evidence, however, is entirely correlational across systems that differ in many confounded dimensions. ComplexDec differs from AudioDec/ScoreDec in input representation (complex STFT vs waveform), encoder/decoder topology (no downsampling/upsampling), quantizer structure (8+8 complex-valued codebooks vs 16 waveform codebooks), training losses (complex MSE/MAE plus multi-resolution mel vs single-resolution mel), and the presence of the SPF (shared only with ScoreDec). DAC differs further in code dimension (1024 vs 256), training-data scale (thousands of hours vs 30 h), and output bandwidth (24 kHz vs 48 kHz). Table I and the Discussion interpret the ordering of out-of-domain degradation as 'high correlation between information loss and domain robustness,' but compression ratio is never manipulated holding other factors fixed. The robust OOD behavior could equally be caused by the complex spectral representation preserving harmonic structure (the Fig. 2 observation), by the absence of temporal downsampling per se, by the higher effective capacity of 256-dimensional codes, or by the SPF interacting with spectral coding, rather than by low information loss as quantified by compression ratio. Thus the empirical system may be valuable, but the paper's stated mechanism is underdetermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ComplexDec, a full-band 48 kHz neural audio codec that operates in the complex STFT domain without temporal downsampling, using 16 10-bit codebooks (8 for the real and 8 for the imaginary part) at a 24 kbps bitrate together with a score-based post-filter. The model is trained on the 30-hour VCTK clean subset and evaluated in-domain and out-of-domain on expressive EARS speech, with AudioDec, ScoreDec, Encodec, and DAC as baselines. The authors report that ComplexDec maintains similar objective and subjective quality in- and out-of-domain whereas the waveform-based baselines degrade markedly, and they attribute this robustness to reduced information loss from temporal and dimensional compression, quantified by a low compression ratio of 1.25 versus 5 for AudioDec.","tokens_in":8837,"tokens_out":10486,"duration_ms":95860,"significance":"If the empirical robustness results hold, the paper makes a useful contribution: it demonstrates that a codec trained on only 30 hours of reading-style speech can generalize to expressive speech, it provides what appears to be the first dedicated out-of-domain robustness evaluation of neural codecs, and it shows that a low-dimensional (256-dim) discrete representation is feasible for full-band coding, which is favorable for regression-based generative models. The whisper-condition observation (AudioDec/ScoreDec degrade less on unvoiced speech) offers a falsifiable nuance consistent with the harmonic-structure account. However, as detailed below, the causal attribution to compression-ratio information loss is not established by the reported comparisons, and several statistical details of the evaluations need strengthening. The contribution is therefore more secure as an empirical demonstration of a robust codec design than as a demonstration of the proposed mechanism.","major_comments":[{"comment":"The compression ratio for ComplexDec is computed as 48000/(150×256)=1.25, but ComplexDec codes both the real and imaginary parts of the 256-bin complex spectrum as two separate 256-dimensional streams (Sec. II.B, II.C). Under the definition given in Sec. II.A, the total latent dimensionality is 512, giving 48000/(150×512)=0.625 rather than 1.25. This discrepancy affects the headline quantitative comparison with AudioDec's ratio of 5 and with DAC's 0.3125. The authors should state explicitly whether H denotes complex coefficients or real values and recompute the ratios consistently; the current figure of 1.25 is not internally consistent with the architecture.","section":"§II.A-II.B, Table I"},{"comment":"The central causal claim that low information loss (as proxied by compression ratio) causes out-of-domain robustness is not isolated by the experiments. ComplexDec differs from AudioDec and ScoreDec in multiple dimensions simultaneously: input representation (complex STFT versus waveform), encoder/decoder topology (no downsampling/upsampling), quantizer structure (8+8 complex codebooks versus 16 waveform codebooks), training losses (complex MSE/MAE plus multi-resolution mel versus single-resolution mel), and the presence of the score-based post-filter (shared only with ScoreDec). DAC differs further in code dimension, training data scale (thousands of hours versus 30 hours), and output bandwidth (24 kHz versus 48 kHz). Table I and the Discussion interpret the observed ordering as 'high correlation between information loss and domain robustness,' but no experiment varies the compression ratio while holding the other factors fixed. The robust OOD behavior could equally be caused by the complex spectral representation preserving harmonic structure (Fig. 2), the absence of temporal downsampling per se, the higher code capacity, or the SPF. I recommend either softening the causal language to 'consistent with' or adding ablations that isolate the compression ratio, e.g., training a waveform-domain ComplexDec variant with the same no-downsampling architecture and 256-dimensional codes, or varying the hop length/number of codebooks within ComplexDec while keeping representation and losses constant.","section":"§II.B, §III.B, §IV"},{"comment":"Objective metrics in Table I are reported as point estimates without error bars or significance tests. Claims that ComplexDec 'achieves similar in-domain and out-of-domain performances' and that baselines 'suffer serious degradation' rely on single numbers, and the metrics are computed after downsampling to 24 kHz (waveform metrics) and 16 kHz (STOI/PESQ), so full-band artifacts are not captured. The authors acknowledge the lack of full-band objective measures, but the downsampled numbers are still used to support the robustness claim. I ask for per-utterance standard deviations, bootstrap confidence intervals, or pairwise significance tests for at least the key comparisons in Table I.","section":"§III.B, Table I"},{"comment":"The MOS evaluation uses only 10 utterances per system and 12 raters, and 95% confidence intervals are reported only for the Overall row. The text uses 'significantly' without any inferential statistics (e.g., paired tests or CI-based comparisons with multiple-comparison correction). For example, in the out-of-domain column, ComplexDec's Overall MOS of 3.99±0.14 and DAC's 3.70±0.16 have nearly overlapping CIs, and in-domain/out-of-domain differences for ComplexDec are not tested. Please provide per-style CIs, per-listener variance, and a justified significance analysis, or restrict the claims to descriptive trends.","section":"§III.C, Table II"}],"minor_comments":[{"comment":"Abstract: 'AuidoDec' should be 'AudioDec'.","section":"Abstract"},{"comment":"Table II title: 'CONFIDENCE INTERNALS' should be 'CONFIDENCE INTERVALS'.","section":"Table II title"},{"comment":"Sec. III.D: 'ComplexDex' should be 'ComplexDec' in the RTF sentence.","section":"Sec. III.D"},{"comment":"Sec. III.B: 'The results demonstrate the high correlation between information loss and domain robustness' implies a quantitative correlation that is never computed; consider replacing with 'are consistent with' or reporting a correlation measure.","section":"Sec. III.B"},{"comment":"Sec. II.B: The phrase 'low spectral dimension setting (e.g. 256-dim)' should clarify that this is the number of complex frequency bins and that real and imaginary streams together constitute 512 real dimensions, to avoid the confusion noted in Major Comment 1.","section":"Sec. II.B"},{"comment":"Sec. III.A: 'The total length of the training data' should be 'The total duration of the training data' for grammatical precision.","section":"Sec. III.A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an empirical systems paper with a plausible but unproven mechanistic explanation. The authors' self-citations to AudioDec and ScoreDec are appropriate given the direct baselines, and I see no citation or scope concerns. The main risk to publication is the gap between the correlation evidence and the causal claim; if the authors are willing to reframe the contribution or add the suggested ablations, the paper would be a strong addition to the audio coding literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the short version: ComplexDec is a genuinely useful empirical contribution, and the paper's central causal story is plausible but not proven by the reported comparisons. Worth engaging with; needs a serious referee.\n\nWhat's new: a full-band 48 kHz neural codec that operates directly on complex STFT spectra, with a dual real/imaginary RVQAE and no temporal downsampling, plus a score-based post-filter. The authors train on 30 h of reading-style VCTK and show it holds up on expressive EARS speech where AudioDec and ScoreDec collapse. The systematic out-of-domain evaluation across several neural codecs is the real contribution; I don't know of another paper in this space that directly tests domain robustness this way. The demo page and training details are a plus; the comparison to open-source Encodec and DAC gives context.\n\nWhat's good: the objective and subjective results line up. ComplexDec's OOD degradation is small, and the subjective table shows it beating Encodec and matching DAC while using far less training data. The spectrogram analysis (AudioDec losing harmonics, ComplexDec preserving them) is a nice qualitative observation. The paper is also honest about some tradeoffs: higher quantization error on high frequencies and slow SPF inference.\n\nSoft spots, in order of importance. First, the causal claim that 'information loss degrades robustness' is not actually isolated. ComplexDec differs from AudioDec/ScoreDec in representation (complex STFT vs waveform), architecture (no down/upsampling), code dimension (256 vs 64), loss functions, and the SPF. Compression ratio is never varied independently, so the robust behavior could come from the spectral representation or the high code dimension rather than low compression per se. The stress-test note gets this right. It doesn't sink the paper, but the intro and conclusion should be softened to 'consistent with' rather than 'caused by.' Second, objective metrics have no error bars and the subjective test is 10 utterances from one dataset with an internal team; the MOS CIs are small but this is still a thin base for strong claims. Third, no code is released, which matters for a reproducibility-oriented paper. Minor: the 'first to explore OOD robustness' claim is stronger than what a quick literature check supports.\n\nBottom line: this is a solid empirical paper with an over-claimed mechanism. For an audio codec venue it deserves peer review, and the authors should be asked to either run an ablative test (e.g., a waveform-based variant with the same code dimension, or a downsampled complex variant) or reframe the mechanism as a design motivation rather than a demonstrated cause. I'd cite the OOD evaluation and the codec design.","headline":"Useful empirical codec paper with an over-claimed causal story; the OOD robustness is real, but the information-loss mechanism is not isolated.","tokens_in":9346,"tokens_out":3947,"would_cite":true,"duration_ms":37417,"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":"ComplexDec claims that coding speech as complex spectra without downsampling preserves naturalness on expressive out-of-domain speech at 24 kbps, even when trained on only 30 hours of reading-style data.","keywords":["neural audio codec","complex spectrum modeling","out-of-domain robustness","information loss","residual vector quantization","score-based post-filter","full-band 48 kHz speech"],"falsifier":"Train ComplexDec with the same complex-spectrum inputs and losses but insert a single downsampling and upsampling stage to raise the compression ratio to about 5; if out-of-domain MOS remains near 3.9 instead of falling toward 1.5, the paper's attribution of robustness to low compression ratio is falsified.","tokens_in":8362,"feed_emoji":"🎧","tokens_out":8451,"duration_ms":76254,"temperature":0.7,"pith_summary":"ComplexDec is a 48 kHz neural audio codec built on a simple hypothesis: the quality collapse that neural codecs show on unfamiliar audio comes from information lost when the encoder throws away temporal resolution and compresses the embedding dimension, not from lack of training data. To test that, the codec operates directly on the real and imaginary parts of short-time Fourier transform spectra at a 150 Hz frame rate with 256 dimensions and no downsampling or upsampling layers, keeping the same 24 kbps bitrate as its waveform-domain baselines. Trained on only about 30 hours of reading-style VCTK speech, it scores nearly the same on in-domain and out-of-domain expressive EARS speech, while AudioDec and ScoreDec drop sharply and open-source Encodec models trained on far more data also degrade. If the hypothesis is right, codec robustness can be bought by preserving information, at a compression ratio near 1.25, rather than by scaling up data, and the resulting low-dimensional complex-spectrum tokens are well suited to downstream generative models.","feed_headline":"Complex spectrum coding keeps speech codec stable on new voices","feed_subtitle":"At 24 kbps, ComplexDec codes loud, whispered, and expressive speech it was never trained on with little quality loss.","key_machinery":"The carrying mechanism is the complex-spectrum RVQAE with no temporal or dimensional compression. Two residual-vector-quantizer autoencoders, one for the real part and one for the imaginary part of the STFT, share the same encoder and decoder but use independent codebooks; the 150 Hz frame rate and 256-dimensional embedding allow the model to keep the 24 kbps bitrate (16 10-bit codebooks, 8 per part) with only a 1.25 compression ratio, versus 5 for AudioDec. A score-based post-filter trained with score matching and reverse-SDE sampling, together with amplitude companding, refines the decoded spectra. The no-downsampling design is what lets the codec carry the long-term harmonic structure that the paper identifies as missing when out-of-domain speech is coded by downsampled baselines.","core_discovery":"On the paper's own terms, the central discovery is that easing information loss in the codec bottleneck transfers to out-of-domain robustness without extra training data. ComplexDec encodes and decodes the real and imaginary components of 48 kHz STFT spectra using two residual-vector-quantizer autoencoders that share an encoder and decoder but keep separate codebooks; with 16 10-bit codebooks (8 per component) at 150 Hz, the bitrate is still 24 kbps but the compression ratio is only 1.25, compared with 5 for AudioDec. In listening tests, overall MOS for ComplexDec is 3.93 in-domain and 3.99 out-of-domain, whereas AudioDec falls from 3.06 to 1.49 and ScoreDec from 4.28 to 2.46; the same pattern appears in waveform and perceptual objective metrics. The paper takes this, plus the harmonic structure visible in spectrograms and the fact that whisper, which lacks harmonics, degrades least, as evidence that lost long-term harmonic information, not data scale, is the root of neural codec domain failure.","pith_inferences":["A cleaner causal test would vary only the compression ratio, for example by changing the code dimension at fixed 150 Hz frame rate, while keeping the complex-spectrum representation, architecture, and losses fixed; the paper does not run that ablation.","The same design rationale predicts that robustness problems will be milder for signals without stable harmonic structure, consistent with the paper's observation that whisper coding degrades least.","Because the perceptual gap between ComplexDec and DAC appears only at full bandwidth, out-of-domain codec evaluation should include 48 kHz listening tests rather than relying on downsampled objective metrics.","If the information-loss mechanism is right, one-step or distilled post-filter inference could make ComplexDec streamable without losing its robustness advantage, since the SPF is the only slow non-causal component."],"forward_implications":["A codec trained on just 30 hours of reading-style speech can keep natural quality on expressive, loud, and whispered speech from a different recording environment, provided it avoids temporal and dimensional compression.","Compression ratio, not just bitrate, is the quantity that predicts out-of-domain robustness: at the same 24 kbps, the lower-ratio codec degrades far less.","A score-based post-filter can sharpen coded spectra but cannot restore harmonics the encoder already lost, so encoder information loss is the bottleneck.","Low-dimensional 256-dimensional complex-spectrum tokens are easier for regression-based generative models to consume than the 1024-dimensional codes used by some low-ratio baselines.","Open-source codecs trained on far more data still show large out-of-domain degradation, which implies that simply scaling training data does not solve the robustness problem."],"supporting_citations":[{"why":"Supplies the AudioDec waveform-domain RVQAE baseline whose out-of-domain degradation is the main comparison.","marker":"[12]"},{"why":"Supplies the ScoreDec baseline and the score-based post-filter method adopted for ComplexDec.","marker":"[18]"},{"why":"Supplies the DAC baseline whose low compression ratio and high-dimensional codes frame the information-loss argument.","marker":"[23]"},{"why":"Supplies the open-source Encodec baselines trained on large data that still degrade out-of-domain.","marker":"[11]"},{"why":"Supplies the VCTK corpus used as the 30-hour reading-style training set.","marker":"[19]"},{"why":"Supplies the EARS expressive-speech corpus used for in-domain and out-of-domain evaluation.","marker":"[20]"},{"why":"Supplies the complex-STFT score-based training and sampling recipe adapted for the post-filter at 48 kHz.","marker":"[26]"}],"fun_headline_variants":["Complex spectra make audio codec robust to unseen voices","Modeling complex spectra boosts codec domain robustness","ComplexDec: robust neural codec with complex spectral modeling","24 kbps codec handles new voices by keeping spectral detail","Less info loss via complex spectra: codec stays stable on new speech"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the out-of-domain gains come from easing information loss by avoiding temporal and dimensional compression, but ComplexDec also changes the input representation, architecture, and training losses at the same time, so that causal link is not isolated.","fun_headline_variants_meta":{"raw":{"variants":["Complex spectra make audio codec robust to unseen voices","Modeling complex spectra boosts codec domain robustness","ComplexDec: robust neural codec with complex spectral modeling","24 kbps codec handles new voices by keeping spectral detail","Less info loss via complex spectra: codec stays stable on new speech"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000686,"raw_usage":{"total_tokens":3087,"prompt_tokens":901,"completion_tokens":2186,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":2104}},"tokens_in":517,"tokens_out":2186,"duration_ms":14241,"temperature":1.0,"reasoning_tokens":2104,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:37:40.087661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ComplexDec with the same complex-spectrum inputs and losses but insert a single downsampling and upsampling stage to raise the compression ratio to about 5; if out-of-domain MOS remains near 3.9 instead of falling toward 1.5, the paper's attribution of robustness to low compression ratio is falsified.","supporting_citations":[{"cited_title":"AudioDec: An open-source streaming high-fidelity neural audio codec,","cited_arxiv_id":null,"evidence_quote":"Supplies the AudioDec waveform-domain RVQAE baseline whose out-of-domain degradation is the main comparison."},{"cited_title":"Score- Dec: A phase-preserving high-fidelity audio codec with a generalized score-based diffusion post-filter,","cited_arxiv_id":null,"evidence_quote":"Supplies the ScoreDec baseline and the score-based post-filter method adopted for ComplexDec."},{"cited_title":"High-fidelity audio compression with improved rvqgan,","cited_arxiv_id":null,"evidence_quote":"Supplies the DAC baseline whose low compression ratio and high-dimensional codes frame the information-loss argument."},{"cited_title":"CSTR VCTK corpus: En- glish multi-speaker corpus for CSTR voice cloning toolkit,","cited_arxiv_id":null,"evidence_quote":"Supplies the VCTK corpus used as the 30-hour reading-style training set."},{"cited_title":"Speech enhancement with score-based generative models in the complex STFT domain,","cited_arxiv_id":null,"evidence_quote":"Supplies the complex-STFT score-based training and sampling recipe adapted for the post-filter at 48 kHz."}],"review_version":1}