{"id":"a66f7a51-542c-4d68-8cfb-dc824baab8de","arxiv_id":"2507.22964","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Polar-ratio features derived from spectral subband centroids, plus a pseudo-F0 feature, lower word error rates and improve gender independence in Vietnamese ASR over a 6-MFCC baseline.","lead":"The authors test new acoustic features for Vietnamese speech recognition that describe how spectral tones move over time, combining them with standard MFCCs. On a small 17-hour Vietnamese corpus, the combined features reduce word errors compared with the 6-MFCC baseline and shrink male-female error gaps, though their advantage over the 13-MFCC baseline is mixed.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The polar angle in the proposed SSCF1/SSCF3–SSCF2/SSCF3 ratio plane is mathematically identical to the raw SSCF1–SSCF2 polar angle, so the claimed 'ratio-plane dynamics' reduce to a radius rescaling by SSCF3 and do not support the paper's central novelty.","rationale":"The paper is a serious empirical feature-engineering study, and I read it in good faith. The experimental setup is nontrivial (hybrid DNN-HMM Kaldi recipe, 7-fold CV, balanced male/female splits), and the Table 3 results are internally consistent in showing that adding the proposed features to the 6-MFCC baseline improves WER in most conditions. My concern targets the specific novelty claimed in Section 3.1. As written, Eqs. (3)-(4) applied to the ratio coordinates produce exactly the same polar angle as the raw SSCF1-SSCF2 plane; the only changed quantity is the radius, scaled by 1/SSCF3. This is an algebraic fact, not a disagreement about consensus. It matters because the abstract and conclusion attribute the gains to characterizing 'acoustic transitions in a ratio plane' and to reducing gender-dependent spectral variation. If the implementation is as written, the ratio-plane angle is not a new dynamic parameter, and the Polar-Ratio comparison in Table 3 tests only a radial rescaling. The reader's weakest assumption (unvalidated transfer from English formant ratios to Vietnamese SSCFs) is related, but my concern is stronger: even if the formant-ratio premise were true, the polar-angle feature cannot benefit from it because the angle is already invariant to the common 1/SSCF3 scaling. I do not recommend rejection because the combined feature set still shows empirical promise and a reformulated contribution (radius normalization plus SSCF0) might survive; but the current manuscript needs either a corrected derivation or an explicit description of a different implementation, plus the proposed ablation, before the central claim can be accepted.","tokens_in":7443,"tokens_out":15837,"duration_ms":186579,"concrete_test":"Reimplement the Polar-Ratio features exactly as specified: compute r1 = SSCF1/SSCF3 and r2 = SSCF2/SSCF3, then apply Eqs. (3)-(4) to get radius and angle; numerically verify on the MICA corpus that angle_ratio equals angle_raw for every frame. Then rerun the 7-fold Kaldi experiments with three input variants: (a) raw Polar features, (b) raw Polar angle plus raw radius divided by SSCF3, and (c) the paper's Polar-Ratio features. If (b) and (c) yield identical WERs, the ratio-plane contribution is fully explained by radial rescaling; if they differ, the exact implementation must be documented so the claimed angle dynamics can be reproduced and evaluated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Section 3.1, the polar parameters are defined by Eqs. (3)-(4) as radius = sqrt(SSCF_{i+1}^2 + SSCF_i^2) and angle = arctan(SSCF_{i+1}/SSCF_i). Substituting the proposed ratio-plane coordinates (SSCF1/SSCF3, SSCF2/SSCF3) gives angle = arctan((SSCF2/SSCF3)/(SSCF1/SSCF3)) = arctan(SSCF2/SSCF1), identical to the angle in the raw SSCF1-SSCF2 plane (for positive SSCF3). The radius becomes sqrt(SSCF1^2 + SSCF2^2)/SSCF3, i.e., the raw radius divided by SSCF3. Therefore the ratio plane adds no new angular/directional dynamics; its only effect is a frame-wise radial rescaling. Consequently, the small and inconsistent differences between the Polar and Polar-Ratio rows in Table 3 (TrainMix 9.89→9.94; TrainMale 11.78→11.64; TrainFemale 11.77→11.53) cannot be evidence for the Peterson-style ratio-invariance mechanism invoked in Section 3.1, especially since [24] already reports gender independence for raw SSCF1-SSCF2 angles. If a different ratio-plane angle was intended (e.g., arctan of normalized delta ratios), the paper does not define it. This does not invalidate the WER improvements of the full feature combination, but it removes the theoretical support for the claimed 'ratio plane' contribution and requires re-analysis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes new acoustic features for Vietnamese ASR: polar-coordinate parameters (radius and angle) computed in a 'ratio plane' of Spectral Subband Centroid Frequencies (SSCF1/SSCF3 and SSCF2/SSCF3), a pseudo-F0 feature based on SSCF0, and combinations of these with MFCCs. The features are evaluated on a 17-hour, 28-speaker Vietnamese corpus using a Kaldi DNN-HMM system with 7-fold cross-validation under TrainMix, TrainMale, and TrainFemale conditions. The paper reports that the proposed feature sets reduce word error rates relative to 6- and 13-dimensional MFCC baselines and exhibit smaller male-female WER gaps in several conditions, e.g., 9.50% versus 12.77% in TrainMix and 11.05% versus 14.06% in TrainMale for the full proposed system.","tokens_in":7780,"tokens_out":3099,"duration_ms":36192,"significance":"If the reported improvements are statistically reliable, the work offers a low-cost, hand-crafted feature alternative for low-resource and tonal language ASR, with the practical advantage of not requiring explicit F0 tracking. The experimental design is transparent: feature equations are explicit, the Kaldi recipe is standard, and the three training/test gender configurations directly address the gender-independence claim. The main weaknesses are that the central theoretical motivation for the ratio plane is undermined by a mathematical identity, the WER comparisons lack significance testing or error bars, and the pseudo-F0 contribution is not validated against measured pitch. These issues make the central claim currently unverified rather than proven.","major_comments":[{"comment":"The proposed polar parameters in the ratio plane (SSCF1/SSCF3, SSCF2/SSCF3) are mathematically equivalent to the raw SSCF1-SSCF2 polar parameters up to a frame-wise radial rescaling. For positive SSCF3, Eq. (4) gives arctan((SSCF2/SSCF3)/(SSCF1/SSCF3)) = arctan(SSCF2/SSCF1), which is exactly the angle in the raw SSCF1-SSCF2 plane, and Eq. (3) gives sqrt(SSCF1^2+SSCF2^2)/SSCF3, i.e., the raw radius divided by SSCF3. Thus the ratio plane introduces no new angular or directional dynamics, and the Peterson-style ratio-invariance mechanism invoked from [19] is not actually implemented in the angle. The small and inconsistent differences between the Polar and Polar-Ratio rows in Table 3 (e.g., TrainMix 9.89 vs. 9.94; TrainMale 11.78 vs. 11.64; TrainFemale 11.77 vs. 11.53) should therefore be attributed to the radial rescaling, not to ratio-based angle invariance. The authors should either correct the derivation, define a genuinely different ratio-plane angle, or reframe the contribution as a radial normalization effect.","section":"Section 5, Table 3"},{"comment":"The paper repeatedly claims that the proposed parameters 'significantly reduce' word error rates and exhibit 'greater gender independence,' but no statistical significance tests, confidence intervals, or per-fold error bars are reported. The 7-fold cross-validation is described, but Table 3 reports only a single WER point per condition, so the reader cannot assess whether the differences (e.g., 9.50% vs. 12.77% in TrainMix) are stable across folds or driven by a particular split. The authors should report per-fold results, standard deviations, and paired significance tests (e.g., matched-pairs or bootstrap) for the key comparisons, especially given the small corpus of 28 speakers and 17 hours.","section":"Section 3.2"},{"comment":"The use of SSCF0 as a pseudo-F0 is asserted without comparing it to a standard F0 estimate or to pitch-tracker output, and without any analysis isolating tonal errors. The conclusion that 'SSCF0 enhances Vietnamese speech recognition by capturing valuable information related to the fundamental frequency' is therefore not directly supported. To justify this claim, the authors should show that SSCF0 tracks F0 on a held-out sample, or present an ablation or error analysis demonstrating that the benefit of adding SSCF0 is concentrated in tonal distinctions rather than in general spectral information.","section":"Section 4.2"}],"minor_comments":[{"comment":"There is a typo: 'sparse datasets' appears as 'spare datasets' in the sentence about low-resource languages.","section":"Section 1"},{"comment":"The statement that 'only the SSCF1-SSCF2 plane was chosen' and that 'other SSCF planes do not improve the recognition results' is not accompanied by any supporting data. Either provide a small ablation table or remove the unsupported claim.","section":"Section 3"},{"comment":"Equation (2) defines the angle from end-to-start differences of a whole transition, while Eqs. (3)-(4) define frame-wise polar parameters from instantaneous SSCF values. The relationship between these two formulations should be clarified, since the frame-wise features are what is used in the ASR experiments.","section":"Section 2"},{"comment":"The phrase 'making the parameters greater independent' should be 'making the parameters more gender-independent' or similar.","section":"Section 5"},{"comment":"The text says in Section 3 that the previous polar parameters 'achieved higher word error rates than the MFCCs' on French data; this phrasing is ambiguous because 'higher' means worse, and the following sentence explains that the polar parameters omit detailed information. Consider rephrasing to clearly state that the earlier features alone were less accurate than MFCCs.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's central novelty rests on the 'ratio plane' concept, but as shown in the report, the polar angle in that plane is identical to the raw angle, and the only change is a radius rescaling. This is a load-bearing mathematical issue that the authors must address. The absence of significance testing is also critical given the very small corpus. If the authors can re-analyze the data to separate the radial-scaling effect from any genuine ratio effect, and add proper statistical evaluation, the work could become a modest but useful contribution. As it stands, the 'ratio-plane' claim is not supported by the equations in the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The empirical result is real, but the headline contribution doesn't survive the math. The polar angle in the proposed 'ratio plane' (SSCF1/SSCF3, SSCF2/SSCF3) is arctan(SSCF2/SSCF1) — exactly the same angle as in the raw SSCF1-SSCF2 plane. The only change is a frame-wise rescaling of the radius by SSCF3. So the Peterson-inspired gender-independence mechanism is not actually a new angular feature; it's a scaling trick. The small differences between the Polar and Polar-Ratio rows in Table 3 (e.g., 9.89% vs 9.94% in TrainMix) are noise, not evidence.\n\nWhat the paper does well is run clean, reproducible experiments. Seven-fold CV on a balanced Vietnamese corpus (28 speakers, 17h), standard Kaldi DNN-HMM with sane hyperparameters, and a clear comparison across train/male/female conditions. The full feature combo — polar parameters plus normalized SSCF0 as pseudo-F0 on top of 6 MFCCs — does reduce WER relative to the 6-MFCC baseline and narrows the male-female gap in cross-gender tests. That result is worth having, even if the ratio-plane framing is wrong.\n\nThe soft spots are the usual ones: no significance testing, no error bars, and a small corpus. 'Significantly' in the abstract and conclusion is not supported by the data. The improvement over the 13-MFCC baseline is only in the cross-gender tasks; in TrainMix the best proposed system (9.50%) is statistically indistinguishable from 13 MFCCs (9.46%). The claim that it 'outperforms' 13 MFCCs is overreach. Also, the choice of the SSCF1-SSCF2 plane was made after seeing which plane worked, which is a selection-on-the-test-set problem. Finally, SSCF0 as pseudo-F0 is an interesting hack, but they never validate it against measured pitch; the MVN normalization may be doing all the work.\n\nWho is this for? People doing feature engineering for low-resource tonal ASR will get a useful data point. The paper deserves a serious referee, but the authors need to either drop the ratio-plane angle claim or re-frame it as a radius-normalization trick, and add significance testing and more careful comparisons before publication.\n\nRecommendation: send to peer review, but expect heavy revision. As is, it's a conditional reject; the empirical core is salvageable.","headline":"The paper has a useful empirical result on Vietnamese ASR, but the claimed 'ratio-plane' angle is mathematically identical to the raw polar angle, so the central novelty is an overclaim.","tokens_in":8399,"tokens_out":4048,"would_cite":true,"duration_ms":42892,"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":"This paper claims that polar-coordinate features computed in SSCF ratio planes, plus a pseudo-F0 from the lowest spectral band, lower Vietnamese ASR word error rates and shrink gender gaps.","keywords":["speech dynamics","acoustic gesture","gender-independent speech recognition","tonal language","low-resource language","spectral subband centroid frequency","polar coordinates","Vietnamese ASR"],"falsifier":"Take the same Vietnamese corpus and measure true formant frequencies and F0; if the SSCF1/SSCF3 and SSCF2/SSCF3 ratios scatter by gender as much as the absolute SSCFs do, or if SSCF0 correlates poorly with measured F0, then the gender-independence claim is not supported and the WER gains could come from added feature dimensions or normalization rather than from the ratio mechanism.","tokens_in":7229,"feed_emoji":"🗣️","tokens_out":8890,"duration_ms":88966,"temperature":0.7,"pith_summary":"This paper argues that dynamic speech information, encoded as polar coordinates in ratio planes of Spectral Subband Centroid Frequencies (SSCFs), makes Vietnamese automatic speech recognition more accurate and less dependent on speaker gender. The authors add these polar-ratio features to a six-dimensional MFCC baseline and use SSCF0, the lowest subband centroid, as a stand-in for pitch to carry tonal information. On a 17-hour Vietnamese corpus, the best configuration lowers word error rate from 12.77% to 9.50% in mixed-gender training and from 14.06% to 11.05% in male-only training, while reaching or beating a 13-MFCC baseline with fewer coefficients. If the claim holds, low-resource tonal ASR can get a cheap, speaker-variation-resistant feature set without explicit pitch tracking.","feed_headline":"Ratio-plane features cut Vietnamese ASR error to 9.5%","feed_subtitle":"Tone-aware pseudo-pitch features beat a 13-MFCC baseline and shrink the male-female accuracy gap.","key_machinery":"The central object is the polar-ratio parameter: for each frame, the SSCFs (spectral subband centroid frequencies, weighted mean frequencies per subband, with SSCF0 tracking the lowest band) are placed in the two ratio planes SSCF1/SSCF3 and SSCF2/SSCF3, and the transition dynamics are summarized by the polar radius $r=\\sqrt{SSCF_{i+1}^2+SSCF_i^2}$ and angle $\\theta=\\arctan(SSCF_{i+1}/SSCF_i)$ computed against the current subband pair. These coordinates avoid the discontinuity of a raw transition angle while capturing the direction and magnitude of spectral movement. The normalized SSCF0, mean-and-variance normalized within each utterance, is appended as a pseudo-F0 for tonal information. The ratio-plane construction carries the argument because it is the mechanism intended to suppress gender-dependent spectral variation.","core_discovery":"The central claim is that the direction and magnitude of spectral transitions, measured as polar radius and angle in the SSCF1/SSCF3 and SSCF2/SSCF3 ratio planes, preserve phonetic information that is relatively invariant across male and female speakers, and that the lowest SSCF subband can serve as a pseudo fundamental frequency for Vietnamese tones. Combined with six MFCCs and their deltas, the proposed parameters reach 9.50% word error in mixed-gender training, 11.05% in male-only training, and 11.01% in female-only training. The gender gap is smaller than the 13-MFCC baseline in all three training conditions and smaller than the six-MFCC baseline in mixed-gender and female-only training. The authors conclude that these dynamic features are a more gender-independent acoustic representation for Vietnamese than absolute-frequency cepstral coefficients.","pith_inferences":["The gender-invariance rationale is borrowed from English formant-ratio data, so the mechanism would be on firmer ground if tested on real Vietnamese formant measurements; the paper does not report such a check.","Because the polar-ratio addition also raises the feature dimension, part of the WER gain could come from the extra dimensions rather than from gender-invariant structure; a control with matched numbers of decorrelated features would isolate the mechanism.","If SSCF0 is a good pitch proxy, the same pseudo-F0 idea may transfer to other tone languages, but only after comparing it against measured F0 on continuous speech."],"forward_implications":["A six-MFCC system fitted with polar-ratio and SSCF0 features can beat a 13-MFCC baseline in cross-gender Vietnamese recognition, suggesting that feature design can substitute for extra filter-bank resolution in low-resource settings.","Training on one gender only generalizes better to the other gender with these features: male-only training drops from 14.06% to 11.05% WER on a mixed test set.","Tonal ASR can avoid fragile explicit pitch extraction if SSCF0 reliably carries pitch-related information.","The same ratio-plane recipe, if the gender-invariance premise holds, is a candidate for other tonal and low-resource languages."],"supporting_citations":[{"why":"Introduces the polar-coordinate parameterization of SSCF transitions that this paper adapts and extends.","marker":"[10]"},{"why":"Shows that angular transitions in SSCF planes are relatively independent of speaker gender and speaking rate, providing the dynamic-plane hypothesis.","marker":"[24]"},{"why":"Supplies the observation that formant ratios F1/F3 and F2/F3 are stable across gender, motivating the ratio plane.","marker":"[19]"},{"why":"Defines SSCFs as formant-like, noise-robust spectral features that can be computed without explicit resonance tracking.","marker":"[18]"},{"why":"Establishes that vocalic transitions are describable by direction and transition rate, the dynamic view underlying the features.","marker":"[4]"},{"why":"Documents the difficulty of robust pitch detection in continuous speech, motivating the pseudo-F0 approach.","marker":"[22]"},{"why":"Provides the mean-variance normalization technique later applied to SSCF0 to reduce speaker variation.","marker":"[12]"},{"why":"Supplies the hybrid DNN-HMM recognition recipe used for all experiments and comparisons.","marker":"[6]"}],"fun_headline_variants":["Gender gap narrows with ratio-plane dynamics for Vietnamese ASR","Pseudo-pitch from subband ratios cuts Vietnamese ASR error to 9.5%","Dynamic spectral transitions yield gender-independent Vietnamese ASR","Ratio-plane features outperform MFCCs in Vietnamese gender gap","SSCF dynamics trim word error and balance male-female ASR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the ratio of two frequency-band locations removes male-female vocal-tract differences for Vietnamese speech just as formant ratios do for English vowels, and that the lowest frequency band can stand in for vocal pitch; neither is checked against direct formant or F0 measurements on this corpus.","fun_headline_variants_meta":{"raw":{"variants":["Gender gap narrows with ratio-plane dynamics for Vietnamese ASR","Pseudo-pitch from subband ratios cuts Vietnamese ASR error to 9.5%","Dynamic spectral transitions yield gender-independent Vietnamese ASR","Ratio-plane features outperform MFCCs in Vietnamese gender gap","SSCF dynamics trim word error and balance male-female ASR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1278,"prompt_tokens":838,"completion_tokens":440,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":349}},"tokens_in":454,"tokens_out":440,"duration_ms":5190,"temperature":1.0,"reasoning_tokens":349,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:36:18.109120+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same Vietnamese corpus and measure true formant frequencies and F0; if the SSCF1/SSCF3 and SSCF2/SSCF3 ratios scatter by gender as much as the absolute SSCFs do, or if SSCF0 correlates poorly with measured F0, then the gender-independence claim is not supported and the WER gains could come from added feature dimensions or normalization rather than from the ratio mechanism.","supporting_citations":[{"cited_title":"Preliminary Study on SSCF-derived Polar Coordinate for ASR","cited_arxiv_id":"2212.01245","evidence_quote":"Introduces the polar-coordinate parameterization of SSCF transitions that this paper adapts and extends."},{"cited_title":"Acoustic gesture modeling","cited_arxiv_id":null,"evidence_quote":"Shows that angular transitions in SSCF planes are relatively independent of speaker gender and speaking rate, providing the dynamic-plane hypothesis."},{"cited_title":"The phonetic value of vowels","cited_arxiv_id":null,"evidence_quote":"Supplies the observation that formant ratios F1/F3 and F2/F3 are stable across gender, motivating the ratio plane."},{"cited_title":"Spectral subband centroid features for speech recognition","cited_arxiv_id":null,"evidence_quote":"Defines SSCFs as formant-like, noise-robust spectral features that can be computed without explicit resonance tracking."},{"cited_title":"Signal dynamics in the production and per- ception of vowels","cited_arxiv_id":null,"evidence_quote":"Establishes that vocalic transitions are describable by direction and transition rate, the dynamic view underlying the features."},{"cited_title":"A com- parative analysis of pitch detection methods under the influence of different noise conditions","cited_arxiv_id":null,"evidence_quote":"Documents the difficulty of robust pitch detection in continuous speech, motivating the pseudo-F0 approach."},{"cited_title":"Efficient cepstral normalization for robust speech recognition","cited_arxiv_id":null,"evidence_quote":"Provides the mean-variance normalization technique later applied to SSCF0 to reduce speaker variation."},{"cited_title":"The kaldi speech recog- nition toolkit","cited_arxiv_id":null,"evidence_quote":"Supplies the hybrid DNN-HMM recognition recipe used for all experiments and comparisons."}],"review_version":1}