{"id":"d9c141f6-523f-4209-9949-bff374625994","arxiv_id":"2506.00861","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"RFA-derived AM/FM rhythm spectrograms, used as handcrafted features or as ViT-BERT inputs, give modest accuracy gains over eGeMAPS and Mel spectrograms on ADReSSo dementia classification, but the reported relative gains are miscomputed.","lead":"This paper uses rhythm formant analysis, which tracks low-frequency amplitude and frequency modulations in speech, to build spectrogram features for detecting dementia and estimating cognitive scores. The authors report small accuracy gains over standard baselines on the ADReSSo challenge, though the claimed relative improvements are inflated by a factor of ten.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority over eGeMAPS is not established: the baseline is cited from [22] rather than re-run in the paper's preprocessing, and the abstract's '14.2%' is a factor-10 inflation of the 1.42% gap shown in Table 3.","rationale":"The reader's weakest_assumption focused on the fixed 100-segment resampling in Section 2, which is a legitimate concern about duration and pause artifacts. However, the more decisive issue is that the headline comparison to eGeMAPS is not a controlled experiment: the baseline number is borrowed from [22] rather than recomputed on the paper's own preprocessed utterances, and the abstract's relative improvements are off by a factor of ten. Even granting the preprocessing, the actual gap in Table 3 is one test-set utterance (47 vs 46 correct) with no significance testing. This does not overturn the conditional verdict, because the paper does provide code and a reproducible pipeline that would allow the required check to be run, and the ViT-BERT Mel baseline appears to have been reproduced in-house. But the central claim as stated should not be accepted until the eGeMAPS baseline is rerun under identical conditions and the comparison is shown to be statistically robust.","tokens_in":7967,"tokens_out":7018,"duration_ms":73117,"concrete_test":"Using the authors' code and ADReSSo, apply their exact Section 2 preprocessing (timestamp-based concatenation and fixed 100-segment windows) to all 237 recordings; extract the 88-dimensional eGeMAPS feature set with openSMILE and train the same SVM with the same 5-fold development hyperparameter search as used for the handcrafted features. Then compare per-utterance predictions on the official 71-utterance test set using McNemar's exact test and a bootstrap 95% confidence interval for the accuracy difference. If the recomputed eGeMAPS accuracy is not below 65.71% with p < 0.05, the handcrafted-feature superiority claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that handcrafted rhythm-spectrogram features outperform eGeMAPS depends on Table 3, where 'eGeMAPS [22]' is a published baseline accuracy, not a result produced by the authors on the same preprocessed utterances. Section 4.1 describes concatenating only subject speech segments via timestamps, and Section 2 imposes a fixed 100-segment resampling; the ADReSSo challenge's eGeMAPS result may have been obtained on raw recordings or with a different feature pipeline and classifier. Without re-running eGeMAPS under identical preprocessing and model selection, the comparison is not controlled. Moreover, the abstract's 'relative improvement of 14.2%' is inconsistent with Table 3: (65.71 - 64.79)/64.79 = 1.42%, and on the 71-utterance test set this is 47 vs 46 correct, a one-sample difference. The same factor-10 error appears for the AM spectrogram versus Mel (13.1% claimed vs 1.31% actual). No significance test or confidence interval is reported, so even the corrected margin is within random variation. The superiority claim should either be supported by a fair, statistically tested baseline or withdrawn.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Rhythm Formant Analysis (RFA) based AM and FM rhythm spectrograms for dementia classification and MMSE regression on the ADReSSo dataset. Two families of features are proposed: handcrafted features (variance of rhythm formant trajectories and 2D-DCT coefficients) fed to SVM/SVR/DT, and data-driven features obtained by replacing Mel spectrograms with rhythm spectrograms in a ViT-BERT fusion system with ASR transcripts. The main claims are that the handcrafted features outperform eGeMAPS by a relative 14.2% in classification accuracy and that AM rhythm spectrograms outperform Mel spectrograms by about 13.1%, with comparable or slightly better regression performance.","tokens_in":8270,"tokens_out":4449,"duration_ms":43076,"significance":"If established, the proposal offers an annotation-free acoustic feature family for dementia assessment, which is practically valuable because RFA avoids manual syllable or word annotation. The paper is clearly structured, releases code, and evaluates on a standard challenge dataset (ADReSSo). The ViT-BERT variant with ASR transcripts also addresses a realistic deployment constraint. However, the headline superiority claims rest on a baseline comparison that is not demonstrated to be controlled, on a one-utterance accuracy gap, and on a factor-10 arithmetic error in the reported relative improvements. With a fair, statistically grounded comparison, the contribution could be a useful addition to speech-based dementia detection, but the current evidence is insufficient to support the abstract's claims.","major_comments":[{"comment":"The relative improvements reported in the abstract and in Section 4.3.1 are inflated by a factor of 10. From Table 3, (65.71 - 64.79) / 64.79 = 1.42%, not 14.2%, for handcrafted features versus eGeMAPS, and (74.29 - 73.33) / 73.33 = 1.31%, not 13.1%, for AM spectrogram versus Mel spectrogram. The sentence in Section 4.3.1 stating 'a relative 13.09% improvement' is similarly inconsistent with the table. These are the paper's central quantitative claims and should be corrected, and any downstream conclusions should be adjusted accordingly.","section":"Abstract and Section 4.3.1 (Table 3)"},{"comment":"The eGeMAPS comparison is not controlled. Table 3 lists 'eGeMAPS [22]' as the baseline, but the paper does not report re-computing eGeMAPS features on the same preprocessed utterances used for the proposed features. Section 4.1 describes isolating only the subject's speech segments via timestamps and concatenating them, and Section 2 imposes a fixed 100-segment representation; the ADReSSo challenge results in [22] may have used different preprocessing, features, and model selection. Without re-running eGeMAPS under identical preprocessing and the same SVM cross-validation scheme, the claimed superiority over eGeMAPS is not established. The same concern applies if the Mel-spectrogram ViT-BERT baseline values in Table 2 are quoted from [11] rather than produced by the authors' pipeline.","section":"Section 4.1 and Table 3"},{"comment":"No significance testing or uncertainty quantification is reported for the handcrafted classification results. On the 71-utterance test set, 65.71% and 64.79% correspond to 47 and 46 correct utterances, respectively, a difference of a single subject. Moreover, the hyperparameter C=3 was selected on the development folds, and only the resulting test accuracy is reported. The authors should report confidence intervals, bootstrap or McNemar significance tests, and ideally the development-fold performance across C values, before claiming that rhythm-spectrogram features outperform eGeMAPS.","section":"Section 4.3.1 and Table 3"},{"comment":"The fixed 100-segment extraction from a 5 s window is a load-bearing preprocessing choice that is not sufficiently justified or controlled. The text says the method 'extract[s] a fixed set of 100 segments from the AM envelope, ensuring consistent temporal segmentation regardless of variations in utterance duration,' which is ambiguous: if short and long utterances are rescaled to the same segment count, the features may encode duration or pause information rather than intrinsic rhythm; if instead a fixed step size is used, shorter utterances will not yield 100 segments. Since dementia speech typically has longer pauses and variable durations, the authors should clarify the resampling procedure, report duration statistics for HC and dementia groups, and show that the classification advantage is not explained by utterance duration or pause fraction.","section":"Section 2"}],"minor_comments":[{"comment":"There is a typo in 'Mel-frequency cepstral co-coefficients' in the first paragraph of the Introduction; it should read 'coefficients.'","section":"Introduction"},{"comment":"The F1-score standard deviation for the Mel baseline is reported as 0.006, while the accuracy standard deviation is 0.67; this disparity is surprising and should be checked for a typo or clarified in the text.","section":"Table 2"},{"comment":"The naming of the spectrogram channels is inconsistent: the text says 'AM and FM spectrograms, along with their ∆ and ∆∆ spectrograms, are each provided as three input channels to ViT,' but Table 2 lists C1, C2, C3 as the channel labels. Please unify the notation.","section":"Section 3.2 and Figure 3"},{"comment":"The regression comparison with eGeMAPS is reported only in prose ('eGeMAPS features achieve a lower RMSE of 6.09, as reported in [22]'); for completeness, this should be included in Table 4 with the same caveat about not being re-run on the authors' preprocessing.","section":"Section 4.3.2"}],"recommendation":"major_revision","confidential_remarks":"The factor-10 percentage error in the abstract is likely a simple arithmetic slip, but it is load-bearing because it is the headline result; it must be corrected in revision. The more serious issue is the uncontrolled eGeMAPS baseline: if the authors cannot re-run eGeMAPS under their exact preprocessing, they should withdraw the superiority claim and present the handcrafted results as comparable or exploratory. The ViT-BERT AM-versus-Mel comparison is more self-contained, but still needs explicit clarification of whether the Mel baseline was re-run in-house and significance testing over the three seeds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is the transfer of RFA rhythm spectrograms to dementia speech. That domain shift is real, and the paper ships code, which is more than many in this area do. The handcrafted feature extraction is coherent, the ViT-BERT adaptation is reasonable, and the regression results are presented honestly enough. I read the stress-test note and largely agree with it.\n\nThe soft spots are the load-bearing ones. The abstract says the handcrafted features beat eGeMAPS by a relative 14.2%, but Table 3 says 65.71 vs 64.79, which is 1.42%. Same for the AM spectrogram versus Mel: 13.1% claimed, 1.31% actual. That is a factor-of-ten inflation, and it needs to be fixed outright. More importantly, the eGeMAPS number is taken from the ADReSSo challenge paper, and the Mel baseline is taken from the ViT-BERT paper; neither baseline was re-run under this paper's preprocessing, which includes concatenating only the subject's speech segments and then forcing a fixed 100-segment representation. That makes the comparison uncontrolled. The accuracy differences are also tiny: on a 71-utterance test set, one or two utterances separate the systems, and no significance test or confidence interval is reported. I would not call this a fatal flaw because the method could still be useful, but the central claim as stated is not established.\n\nThe fixed 100-segment resampling is a genuine worry too. Dementia speech has variable durations and long pauses. If the resampling rescues duration information or the pause pattern differently for short and long recordings, the classifier could be picking up recording length rather than rhythm formants. The paper does not address this.\n\nWho is this for? Speech pathologists and dementia researchers who want a quick, annotation-free acoustic feature. It is not a breakthrough, and the modest gains might evaporate under a fair baseline, but the idea is worth testing. A serious referee should see this manuscript, but the authors must correct the arithmetic, re-run the baselines under identical preprocessing, and add at least some uncertainty estimate. Without those, the paper should not be accepted as is.","headline":"Interesting domain application with a clear pipeline and released code, but the central claim is weakened by a decimal-point error in the reported improvements and by baselines that are cited rather than reproduced.","tokens_in":8853,"tokens_out":2364,"would_cite":false,"duration_ms":25390,"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":"AM rhythm spectrograms outperform Mel spectrograms for dementia classification.","keywords":["dementia detection","rhythm formant analysis","rhythm spectrograms","amplitude modulation","frequency modulation","vision transformer","BERT","MMSE estimation"],"falsifier":"Take the ADReSSo test set, measure total utterance duration, and retrain the handcrafted SVM with duration and pause proportion added as control features, or match the healthy and dementia groups on duration; if the 14.2% accuracy advantage over eGeMAPS disappears or falls below significance, the central claim is falsified.","tokens_in":7794,"feed_emoji":"🧠","tokens_out":4956,"duration_ms":42711,"temperature":0.7,"pith_summary":"This paper tries to establish that rhythm spectrograms derived from Rhythm Formant Analysis (RFA) capture dementia-related changes in the low-frequency amplitude and frequency modulations of spontaneous speech, and that these annotation-free features beat standard acoustic baselines for classifying dementia and estimating MMSE. On the ADReSSo challenge data, handcrafted rhythm-spectrogram features reach 65.71% accuracy, a relative 14.2% improvement over eGeMAPS, and AM rhythm spectrograms in a vision-transformer/BERT fusion reach 74.29% accuracy, about 13.1% relative improvement over Mel spectrograms. If true, RFA offers a cheap, fully automatic speech biomarker for cognitive screening that does not require manual transcripts or forced alignment.","feed_headline":"AM rhythm spectrograms beat Mel features for dementia detection","feed_subtitle":"Speech-only dementia tests: rhythm features beat standard acoustic baselines by 14.2% and 13.1%.","key_machinery":"The central object is the rhythm spectrogram, a time-frequency representation built from the low-frequency (<10 Hz) spectrum of the amplitude-modulation (AM) envelope and the F0-based frequency-modulation (FM) envelope of speech. Speech is split into a fixed grid of 100 overlapping 5-second segments, each segment's low-frequency spectrum is computed, and the spectra are stacked in time to form the spectrogram; rhythm formants are spectral peaks tracked across time, and their variance plus low-order 2D-DCT coefficients form the handcrafted features. In the data-driven path, the spectrograms with their delta and double-delta versions are fed as three channels to a pretrained vision transformer (ViT), whose output is concatenated with BERT embeddings of Wav2Vec2.0 ASR transcripts and passed through a single trainable fully connected layer. The machinery works by converting slow prosodic and articulatory variation, previously measured through manual syllable or word annotation, into an image-like representation that can be processed automatically.","core_discovery":"On its own terms, the discovery is that the long-term temporal modulations of the speech signal below 10 Hz, captured as AM and FM rhythm spectrograms, carry discriminating information about dementia. The authors show this information can be exploited two ways: as handcrafted variance and 2D-DCT features with an SVM, and as three-channel image inputs to a vision transformer whose embeddings are fused with BERT-based linguistic embeddings from ASR transcripts. In both schemes the rhythm-spectrogram features outperform the corresponding baselines (eGeMAPS for the handcrafted path, Mel spectrograms for the fusion path), while giving comparable MMSE regression. The authors claim this demonstrates RFA as a useful annotation-free alternative for dementia assessment.","pith_inferences":["Editorial inference: because every utterance is squeezed into a fixed 100-segment grid, short and long recordings are rescaled differently; a duration-matched control would reveal whether part of the accuracy gain reflects recording length rather than genuine rhythm.","Editorial inference: the 13.1% AM-over-Mel margin is specific to this dataset and fusion configuration, so on other dementia corpora or with different pretrained transformers the ranking could shift.","Editorial inference: a natural next test is whether rhythm spectrograms generalize to mild cognitive impairment or frontotemporal dementia, where prosodic breakdown differs from Alzheimer's dementia.","Editorial inference: the two handcrafted feature families capture different aspects (variance yields better F1, DCT better accuracy), so a learned weighting of both could push performance further."],"forward_implications":["Handcrafted rhythm-spectrogram features could make dementia screening deployable on lightweight devices, since they require only an SVM rather than a large neural network.","The AM spectrogram's superiority over Mel suggests slow amplitude-modulation dynamics are a stronger acoustic marker for dementia than static spectral shape.","Using Wav2Vec2.0 ASR transcripts instead of manual transcriptions removes the need for expert transcription and makes the pipeline usable at scale.","Comparable MMSE regression with rhythm features offers a non-invasive route to estimating cognitive severity from speech.","Because RFA needs no syllable or word segmentation, the same pipeline can be applied to any spontaneous speech recording without annotation."],"supporting_citations":[{"why":"Supplies the Rhythm Formant Analysis method and the AM/FM rhythm spectrogram computation that the paper adapts.","marker":"[15]"},{"why":"Supplies the ViT-BERT end-to-end fusion architecture and the Mel-spectrogram baseline results that the data-driven path is compared against.","marker":"[11]"},{"why":"Supplies the ADReSSo dataset and the eGeMAPS baseline accuracy and MMSE regression results used for comparison.","marker":"[22]"},{"why":"Supplies the RAPT pitch-tracking algorithm used to compute the F0 contour that forms the FM envelope.","marker":"[20]"},{"why":"Supplies the Wav2Vec2.0 self-supervised ASR model that generates transcripts for the BERT branch.","marker":"[28]"},{"why":"Supplies the Geneva Minimalistic Acoustic Parameter Set (GeMAPS) used as the handcrafted-feature baseline.","marker":"[29]"},{"why":"Supplies the peak-picking algorithm used to locate rhythm formants in each low-frequency spectral slice.","marker":"[23]"}],"fun_headline_variants":["AM/FM rhythm spectrograms beat Mel for dementia detection","Handcrafted rhythm features top eGeMAPs by 14.2% in dementia","RFA spectrograms surpass Mel for dementia classification","Fusion of rhythm spectrograms and BERT lifts dementia scores"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that squeezing every utterance into the same fixed 100 five-second segments leaves the rhythm content comparable across speakers; if recording length or pause distribution differs systematically between people with and without dementia, the reported gains could come from duration rather than from rhythm.","fun_headline_variants_meta":{"raw":{"variants":["AM/FM rhythm spectrograms beat Mel for dementia detection","Handcrafted rhythm features top eGeMAPs by 14.2% in dementia","RFA spectrograms surpass Mel for dementia classification","Fusion of rhythm spectrograms and BERT lifts dementia scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1554,"prompt_tokens":849,"completion_tokens":705,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":629}},"tokens_in":465,"tokens_out":705,"duration_ms":6799,"temperature":1.0,"reasoning_tokens":629,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:55:41.626698+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the ADReSSo test set, measure total utterance duration, and retrain the handcrafted SVM with duration and pause proportion added as control features, or match the healthy and dementia groups on duration; if the 14.2% accuracy advantage over eGeMAPS disappears or falls below significance, the central claim is falsified.","supporting_citations":[{"cited_title":"A two- step attention-based feature combination cross-attention system for speech-based dementia detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the Rhythm Formant Analysis method and the AM/FM rhythm spectrogram computation that the paper adapts."},{"cited_title":"Evaluation of speech-based pro- tocol for detection of early-stage dementia","cited_arxiv_id":null,"evidence_quote":"Supplies the ViT-BERT end-to-end fusion architecture and the Mel-spectrogram baseline results that the data-driven path is compared against."},{"cited_title":"The rhythms of rhythm,","cited_arxiv_id":null,"evidence_quote":"Supplies the ADReSSo dataset and the eGeMAPS baseline accuracy and MMSE regression results used for comparison."},{"cited_title":"Prosody analysis as a tool for differential diagnosis of cognitive impairment,","cited_arxiv_id":null,"evidence_quote":"Supplies the RAPT pitch-tracking algorithm used to compute the F0 contour that forms the FM envelope."},{"cited_title":"A robust algorithm for pitch tracking (rapt),","cited_arxiv_id":null,"evidence_quote":"Supplies the Wav2Vec2.0 self-supervised ASR model that generates transcripts for the BERT branch."},{"cited_title":"r9y9/pysptk: 0.1. 14,","cited_arxiv_id":null,"evidence_quote":"Supplies the Geneva Minimalistic Acoustic Parameter Set (GeMAPS) used as the handcrafted-feature baseline."},{"cited_title":"The Future of Prosody: It’s about Time,","cited_arxiv_id":null,"evidence_quote":"Supplies the peak-picking algorithm used to locate rhythm formants in each low-frequency spectral slice."}],"review_version":1}