{"id":"582a8c1b-74ed-4b5d-a834-2baa727b5bc7","arxiv_id":"2505.19626","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"EEG decoding of Mandarin pitch is more accurate for speaker-normalized than raw pitch across multiple speakers, suggesting the brain encodes relative, speaker-independent pitch.","lead":"This paper asks whether the brain tracks the shape of a speaker's pitch rather than raw pitch height, using brain recordings while people hear Mandarin syllables. A new neural network called CE-ViViT decoded pitch contours from EEG and did better on speaker-normalized pitch, suggesting the brain stores relative pitch.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported normalized-pitch decoding advantage may be an artifact: normalized outputs are rescaled with ground-truth per-speaker mean/variance, so even a zero-information model predicts speaker means.","rationale":"The paper's central claim rests entirely on the raw-versus-normalized decoding comparison in Table 2. The comparison is biased by the evaluation protocol: normalized predictions are back-transformed to Hz using ground-truth per-speaker mean and variance, giving the normalized decoder oracle access to speaker statistics that the raw decoder is not given. A zero-information normalized model then predicts each trial's speaker mean, and its MAE is smaller than a zero-information raw model's MAE by exactly the between-speaker F0 variability. The absence of an effect in single-speaker session 1 is the pattern the artifact predicts. Even if one grants that the z-score transform is the brain's normalization, the current experiment cannot distinguish neural encoding of relative pitch from this rescaling artifact. The reader's weakest assumption (that z-score is the neural normalization) is secondary: the more immediate problem is that the evaluation itself is not fair. The proposed concrete test with a trivial speaker-mean baseline would settle whether the reported effect survives; until then the headline conclusion is unsupported. The model and dataset may be useful, but the central neuroscientific claim should not be accepted on this evidence.","tokens_in":7812,"tokens_out":13744,"duration_ms":160757,"concrete_test":"Run a no-information baseline on session 2: for every trial, output 0 as the normalized pitch contour, then rescale using the true per-speaker mean and standard deviation exactly as in Section 4.1, and compute MAE/RMAE/GPE. Compare against the 'Proposed norm' row of Table 2. Also compute the raw baseline that outputs the global mean F0 for every trial. If the trivial normalized baseline achieves MAE close to or below 30.5 Hz, or if the gap between the two trivial baselines accounts for most of the reported raw-versus-norm gap, the central claim is an artifact. As a secondary check, rescale normalized predictions using speaker statistics estimated from the training set only.","verdict_should_be":"REJECT","load_bearing_attack":"In Section 4.1, speaker-normalized predictions are rescaled to the raw F0 scale using the speaker's mean and variance before computing MAE/RMAE/GPE, while raw predictions are evaluated directly. This injects oracle per-speaker statistics into the normalized condition. A trivial decoder that always outputs the z-score mean (0) becomes, after rescaling, a per-trial prediction of that speaker's mean F0, with MAE = E|F0 - μ_s|. The corresponding trivial raw baseline (global mean) has MAE = E|F0 - μ_g|; the gap between these baselines is between-speaker F0 variability and is present even if EEG contains no pitch information. The model's optional identity vector does not remove this asymmetry: the raw condition must learn speaker means from data, while the normalized condition is given them for free during metric computation. The reported normalized advantage in session 2 (~15-18 Hz MAE) is of that magnitude, and session 1 shows no advantage because only one speaker is present - exactly the artifact pattern. No trivial-speaker-mean baseline or training-only speaker-statistic rescaling is reported, so Table 2 does not establish that EEG decodes speaker-normalized relative pitch.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a new EEG dataset of Mandarin monosyllable perception (29 listeners, two stimulus sessions: one speaker and six speakers) and proposes a convolutional embedding plus Video Vision Transformer model (CE-ViViT) to regress fundamental-frequency (F0) contours from EEG. The authors compare decoding of raw F0 with decoding of per-speaker z-score normalized F0, reporting that normalized pitch is decoded more accurately than raw pitch in the six-speaker session (32.7% average MAE improvement) while no clear advantage appears in the single-speaker session. From this they conclude that Mandarin tonal perception relies on speaker-normalized relative pitch and that normalization occurs at the phoneme level. The model is also ablated to justify its components.","tokens_in":8097,"tokens_out":3555,"duration_ms":36294,"significance":"If the central claim were established, the paper would extend invasive ECoG evidence of speaker-normalized relative pitch encoding (Tang et al., 2017) to non-invasive EEG and to isolated monosyllables, which is a meaningful contribution for both neuroscience and auditory EEG decoding. The paper's strengths are the collection of a new auditory EEG dataset with controlled speaker/tone/phoneme variation, the use of multiple SOTA baseline models, and a clearly described architecture with an ablation study. However, the main conclusion currently rests on a metric-comparison asymmetry and on aggregate numbers without variance or significance testing; the significance of the paper is therefore contingent on a careful re-analysis.","major_comments":[{"comment":"The central comparison is not fair as implemented. The text states that speaker-normalized pitch contours are 'rescaled to the original raw pitch contours’ scale using the speaker’s mean and variance during metric calculation', while raw predictions are evaluated directly. This injects oracle per-speaker statistics into the normalized condition. A trivial normalized decoder that always outputs the z-score mean (0) becomes, after rescaling, a per-trial prediction of that speaker's mean F0, with MAE = E|F0 − μ_s|; the corresponding raw trivial baseline (global mean) has MAE = E|F0 − μ_g|. The gap between these baselines is largely between-speaker F0 variability and exists even if the EEG contains no pitch information. This explains the observed pattern: no advantage in session 1 (one speaker) and a large advantage in session 2 (six speakers). The optional identity vector does not remove this asymmetry, because the raw condition must learn speaker means from data while the normalized condition is given them for free during metric computation. The authors should report trivial speaker-mean and global-mean baselines, and/or rescale both conditions using training-set speaker statistics, and/or evaluate all predictions on a common normalized scale.","section":"Section 4.1 and Section 3.3"},{"comment":"No variance, confidence interval, or significance test is reported for any table, yet the conclusion depends on 'no significant difference' in session 1 and a 32.7% average improvement in session 2. With only 3 target subjects, the reported aggregate MAE values could be driven by one subject or by a small number of trials. The test-split description is also ambiguous: the text says 26 participants were used in training and 3 were target subjects, but then says the dataset was partitioned 80/20 into training and validation. The authors should clarify whether target subjects' data were held out entirely, report per-subject results, and provide paired significance tests (or effect sizes with confidence intervals) for the raw-versus-normalized comparison.","section":"Tables 1 and 2, Section 3.2"},{"comment":"The assumption that per-speaker z-score normalization of F0 is the brain's speaker normalization is not tested. If the auditory system normalizes using a different transform (e.g., log-F0, percentile ranks, or an estimated speaker range), then the 'speaker-normalized' target is not the hypothesized neural code, and the decoding comparison does not test relative-pitch encoding. The authors should either include alternative normalization schemes and show the decoding advantage is specific to their chosen transform, or benchmark their target against an independent neural-encoding model. This is a correctness-risk issue rather than a circularity, but it is load-bearing for the interpretation in Section 4.1.","section":"Section 3.1 and Section 4.1"}],"minor_comments":[{"comment":"There are typos and wording issues: 'poorly undared' should be 'poorly understood'; 'as showed' should be 'as shown'; 'V ocoder' should be 'vocoder'.","section":"Introduction"},{"comment":"The notation is inconsistent: the model input is defined as X in R^{C×T}, but the preprocessing text says data were standardized into matrices of dimensions (500, 60). Please clarify which dimension is time and which is channels, and note that the 64-electrode cap results in 60 channels after some processing.","section":"Section 3.1"},{"comment":"The session description is ambiguous: 'each speaker producing 20 monosyllabic sounds across all four tones (80 tokens per speaker)' could mean 20 syllables per tone or 20 syllables total; please state the design explicitly.","section":"Section 3.1"},{"comment":"The paper does not mention ethics approval, informed consent, or data/code availability. For a human EEG study, the ethics statement should be included; the dataset and code availability also need to be stated.","section":"Section 3.2"},{"comment":"Figure 2 shows example reconstructions, but there is no quantitative description of how representative these examples are; consider adding a panel with error bounds or a side-by-side failure case.","section":"Section 4.1 and Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's core finding would be interesting if supported, but the current evaluation asymmetry is severe enough that the main conclusion cannot be accepted as it stands. The recommended revision is substantive but feasible: add trivial baselines, symmetrize the evaluation, report per-subject and significance statistics, and address the normalization-transform assumption. The manuscript would also benefit from a statement on data availability and ethics approval, which is standard for this type of study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper asks a good question: does the auditory cortex encode speaker-normalized relative pitch at the phoneme level, and can EEG decode that representation? The experimental design is genuinely new — Mandarin monosyllables across six speakers, with raw vs. z-scored F0 targets, and a new architecture (CE-ViViT) that does fine as a regression model. I credit the work for posing the question cleanly and for including a single-speaker control that shows no normalization advantage, which is the right sanity check.\n\nThat said, the central result in Table 2 is not yet established. The evaluation rescales normalized predictions to raw F0 using the speaker's mean and variance before computing MAE/RMAE/GPE, while raw predictions are scored directly. That asymmetry hands the normalized condition the per-speaker mean for free. A decoder that outputs the z-score mean (0) becomes, after rescaling, a per-trial prediction of that speaker's mean F0, with error equal to within-speaker F0 deviations. The raw condition has to learn speaker means from data. The gap between those two trivial baselines is between-speaker F0 variability, and it is present even if EEG contains no pitch information at all. The reported session 2 advantage (~15-18 Hz MAE) is exactly that magnitude, and session 1 showing no difference is exactly the artifact pattern. The stress-test note is right, and I checked the paper's own text: Section 4.1 states the rescaling explicitly, and no trivial-speaker-mean baseline is reported. The authors needed to report a no-information decoder baseline, or evaluate normalized predictions in z-space with a corresponding z-space raw baseline, or withhold speaker statistics from the metric until after training. Without that, the interpretation 'the brain encodes relative pitch' is not supported by the comparison, even though the hypothesis itself remains plausible and is consistent with the prior ECoG work [7].\n\nIn smaller key: no variance or significance testing anywhere, only three target subjects, and the test split is described as a validation set without a clear held-out test protocol. These are fixable but they matter for a claim that rests on a null result in session 1 and a 32.7% difference in session 2.\n\nWho is this for? Speech neuroscientists and EEG decoding folks will want to read it for the dataset and the architecture, not yet for the neural conclusion. It deserves a serious referee: the question is important, the prior work is cited correctly, and the artifact is fixable with baselines and a cleaner evaluation. I would not cite the relative-pitch claim in its current form; I would cite the dataset/model once code and data are out. Recommend conditional acceptance with mandatory significance testing, subject-level results, and a trivial-speaker-mean baseline.\n\nBest.","headline":"The neural claim is plausible but the decoding comparison is not yet convincing: the normalized condition may be handed speaker means for free at evaluation time, so the 32.7% advantage is not established.","tokens_in":8605,"tokens_out":689,"would_cite":false,"duration_ms":8599,"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":"Mandarin speech perception encodes pitch relative to the speaker, and the normalization happens within an isolated syllable.","keywords":["EEG decoding","Mandarin tones","speaker normalization","relative pitch","pitch contour","CE-ViViT","speech perception","fundamental frequency"],"falsifier":"A concrete result that would settle the claim would be a multi-speaker decoding study in which raw-pitch targets generalize to held-out speakers as well as normalized-pitch targets do, or in which the normalized-pitch advantage disappears when the z-score is replaced by a different per-speaker transform such as log-frequency centering. Either outcome would undermine the conclusion that the brain encodes speaker-normalized relative pitch at the phoneme level.","tokens_in":7653,"feed_emoji":"🧠","tokens_out":6042,"duration_ms":53499,"temperature":0.7,"pith_summary":"Using a new EEG dataset in which listeners heard Mandarin monosyllables from one speaker and from six speakers, this paper asks whether the brain tracks absolute pitch contours or pitch contours normalized to each speaker's range. A regression model called CE-ViViT reconstructed pitch contours directly from the EEG, and across all tested systems the speaker-normalized contours decoded more accurately whenever multiple speakers were present, with an average MAE improvement of 32.7%. On the single-speaker session normalized and raw pitch decoded equally well, which the authors use to rule out the possibility that the gain comes only from a smaller dynamic range. The paper concludes that Mandarin perception relies on speaker-normalized relative pitch rather than absolute pitch, and that the normalization happens at the phoneme level during tonal perception.","feed_headline":"EEG decoding suggests Mandarin pitch is speaker-relative","feed_subtitle":"Six speakers: normalized pitch contours decode from EEG better than raw pitch, evidence of phoneme-level normalization.","key_machinery":"The object that carries the argument is the speaker-normalized pitch contour, defined here as the fundamental frequency of each syllable converted to a z-score with that speaker's mean and standard deviation. The tool built to read this contour out of brain activity is CE-ViViT, a regression model that embeds raw EEG via parallel temporal, dilation, and channel convolutional blocks refined by ScConv, fuses the branches with DANE, encodes global relations with a two-stage Video Vision Transformer, and maps the representation to the F0 curve through a pitch head trained with mean squared error. Comparing how accurately normalized versus raw contours are decoded in a one-speaker and a six-speaker design is the experiment that does the work.","core_discovery":"The paper's central claim is that when a listener hears a Mandarin syllable, the auditory system encodes the pitch contour relative to the speaker's own pitch range, and this normalization is complete at the phoneme level, before any sentence context is available. The evidence is a decoding contrast: on the six-speaker session, the proposed CE-ViViT and the three comparison systems all reconstructed per-speaker z-scored pitch contours with lower error than raw F0 contours, while on the single-speaker session the two targets were decoded equally well. Because the single-speaker result shows that normalization of the target alone does not help, the multi-speaker advantage is interpreted as neural encoding of relative pitch. The authors state the conclusion directly: Mandarin perception relies on speaker-normalized relative pitch rather than absolute pitch, and the normalization process occurs at the phoneme level during tonal perception.","pith_inferences":["The authors do not test whether the brain's transform is exactly a z-score; identical decoding comparisons using other per-speaker transformations, such as rank normalization or log-frequency centering, would show which code the neural response actually matches.","If the relative-pitch claim is right, a decoder trained on several speakers should transfer normalized-pitch decoding to a new speaker better than it transfers raw-pitch decoding, which is a testable extension of the reported result.","The monosyllabic design leaves open whether sentence context adjusts the normalization; the authors flag sentence-level context as future work, and one would predict that context either sharpens or overrides the phoneme-level normalization.","The result suggests hearing devices for tonal languages could preserve speaker-relative contour shape rather than exact frequencies, since the brain appears to read pitch that way."],"forward_implications":["Decoders for tonal speech brain-computer interfaces should target speaker-relative pitch contours, because that is the representation the EEG signal appears to carry.","Accounts of tone perception that treat absolute F0 as the perceptual feature omit the normalization step that the decoding results place inside the phoneme.","Single-speaker EEG studies cannot separate absolute from relative pitch encoding, so multi-speaker designs are necessary for this question.","The phoneme-level finding means speaker normalization does not need sentence context, which constrains where in the auditory pathway it could occur.","The average 32.7% MAE improvement on the six-speaker session gives later studies a concrete benchmark for detecting speaker-normalized pitch in non-invasive recordings."],"supporting_citations":[{"why":"Provides the intracranial finding that auditory cortex encodes speaker-normalized relative pitch, the hypothesis this EEG study extends.","marker":"[7]"},{"why":"Supplies behavioral evidence that listeners perceive pitch location within a speaker's F0 range, motivating the normalization hypothesis.","marker":"[1]"},{"why":"Provides the Tone Perfect monosyllabic stimuli used for both the single-speaker and six-speaker sessions.","marker":"[25]"},{"why":"Serves as an EEG speech-envelope regression baseline whose normalized-versus-raw comparison reproduces the paper's effect.","marker":"[16]"},{"why":"Serves as a second baseline system whose multi-speaker results also favor speaker-normalized pitch.","marker":"[17]"},{"why":"Serves as a third baseline, an adapted WaveNet decoder, that shows the same normalized-pitch advantage.","marker":"[29]"},{"why":"Provides prior EEG classification results for Mandarin vowels and tones that the paper builds on as non-invasive decoding context.","marker":"[15]"}],"fun_headline_variants":["Mandarin pitch decoded as speaker-relative from EEG","EEG reveals Mandarin pitch is speaker-normalized","Brain encodes Mandarin tones relative to speaker","Phoneme-level pitch normalization seen in EEG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The interpretation depends on the assumption that the brain's speaker normalization is exactly the per-speaker z-score of fundamental frequency; if the brain normalizes pitch by some other rule, the improved decoding of z-scored contours would not demonstrate relative-pitch encoding.","fun_headline_variants_meta":{"raw":{"variants":["Mandarin pitch decoded as speaker-relative from EEG","EEG reveals Mandarin pitch is speaker-normalized","Brain encodes Mandarin tones relative to speaker","Phoneme-level pitch normalization seen in EEG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2876,"prompt_tokens":826,"completion_tokens":2050,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":2001}},"tokens_in":442,"tokens_out":2050,"duration_ms":13875,"temperature":1.0,"reasoning_tokens":2001,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:09:38.627035+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete result that would settle the claim would be a multi-speaker decoding study in which raw-pitch targets generalize to held-out speakers as well as normalized-pitch targets do, or in which the normalized-pitch advantage disappears when the z-score is replaced by a different per-speaker transform such as log-frequency centering. Either outcome would undermine the conclusion that the brain encodes speaker-normalized relative pitch at the phoneme level.","supporting_citations":[{"cited_title":"Perception of pitch location within a speaker’s f0 range,","cited_arxiv_id":null,"evidence_quote":"Provides the intracranial finding that auditory cortex encodes speaker-normalized relative pitch, the hypothesis this EEG study extends."},{"cited_title":"Decoding Speaker-Normalized Pitch from EEG for Mandarin Perception","cited_arxiv_id":"2505.19626","evidence_quote":"Supplies behavioral evidence that listeners perceive pitch location within a speaker's F0 range, motivating the normalization hypothesis."},{"cited_title":"Magnetoencephalography—theory, instru- mentation, and applications to noninvasive studies of the working human brain,","cited_arxiv_id":null,"evidence_quote":"Provides the Tone Perfect monosyllabic stimuli used for both the single-speaker and six-speaker sessions."},{"cited_title":"Synthesiz- ing speech from ecog with a combination of transformer-based encoder and neural vocoder,","cited_arxiv_id":null,"evidence_quote":"Serves as an EEG speech-envelope regression baseline whose normalized-versus-raw comparison reproduces the paper's effect."},{"cited_title":"Decoding speech perception from non-invasive brain record- ings,","cited_arxiv_id":null,"evidence_quote":"Serves as a second baseline system whose multi-speaker results also favor speaker-normalized pitch."},{"cited_title":"Decoding and synthesizing tonal language speech from brain activity,","cited_arxiv_id":null,"evidence_quote":"Provides prior EEG classification results for Mandarin vowels and tones that the paper builds on as non-invasive decoding context."}],"review_version":1}