{"id":"97041e39-8e34-425f-bb3c-ef0b7b3e1617","arxiv_id":"1908.07409","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A rhythm-only, onset-based Query by Humming system is proposed and tested on three songs, claiming better speed, memory, and accuracy than pitch-based methods without a direct comparison.","lead":"The paper builds a Query by Humming system that identifies songs using only the timing of note onsets, skipping pitch entirely. It derives three onset detectors and tests them on three hummed queries against a ten-song database.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical superiority claim is untested: 3 hummed queries, 10-song database, no QBH baseline, and same-verse query assumption.","rationale":"I read the paper as a proof-of-concept; the statistical derivations around GLRT and chi-square power bounds are generally coherent and the experimental code is provided. The central claim, however, is comparative: onset-only QBH is claimed to be better than existing methods in speed, memory, and empirically in accuracy. That claim is load-bearing because the whole contribution is the claimed advantage of dropping pitch information. The reported experiments cannot carry that weight: 3 queries, a 10-song database with the target songs included, no baseline, no runtime or memory measurements, and hyperparameters selected on the first song. The same-verse/linear-rescaling assumption is a further restriction that the paper itself concedes in Section 7. The reader's conditional verdict already captures this uncertainty. I do not see an internal inconsistency that would force rejection, but the acceptance conditions should explicitly require a baseline comparison and a stress test of the same-verse assumption.","tokens_in":18365,"tokens_out":7189,"duration_ms":75703,"concrete_test":"On the same 10-song database and the same 3 hummed queries (ideally 30 or more), run a minimal pitch-contour QBH baseline using DTW with the same rank-1 metric, and measure wall-clock and memory usage. If the onset-only system does not match or beat this baseline on rank-1 accuracy and reported resources, the abstract's comparative claim fails. To control for the same-verse assumption, also run queries that start at a later note or omit a note; if rank drops, the Section 7 limitation is not cosmetic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that an onset-only QBH system is \"better than existing methods... empirically in terms of accuracy.\" The experimental section provides no basis for this comparison. Section 5 reports three hummed queries against a 10-song database that includes the query songs; no existing QBH or pitch-based system is implemented, and the references cite no QBH literature. On a 10-song database, even a trivial fixed ranking yields 10% accuracy, so 3 queries cannot distinguish the method from chance, let alone establish superiority. The pipeline also relies on the Section 4.1.2 assumption that the hummed segment is the same verse and that a linear tempo rescaling aligns onsets one-to-one; Section 7 admits the system cannot handle a user who does not know where the first verse ends. Under missed onsets, the Jingle Bells run itself fails: Energy and Spectral Dissimilarity rank the wrong song first (Table 3), so reported accuracy is tied to one detector and to hyperparameters tuned on the first song. Thus the correctness risk is that the comparative claim is an artifact of the evaluation setup rather than a property of the method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an onset-only Query by Humming (QBH) system. It derives three onset-detection functions from GLRT principles: a local energy detector, a spectral dissimilarity detector, and a dominant spectral dissimilarity detector (Sections 3.1-3.3). Peak detection is performed by threshold and local-maximum comparison (Section 3.4), and probability bounds for detection and false positives are derived via Boole-Frechet inequalities (Section 3.5). Retrieval is done by subset matching and correlative matching with a linear tempo-scaling search (Sections 4.1-4.2). The experiments use three hummed queries against a 10-song database (Section 5), with claim of perfect retrieval for two songs and mixed results for the third, plus power calculations based on estimated noise variance and SSNR values.","tokens_in":18590,"tokens_out":4202,"duration_ms":41435,"significance":"If the central claim were established, the paper would be practically significant: an onset-only representation would reduce storage and search cost relative to pitch- or audio-based QBH, and might be robust to off-key humming. The statistical derivations in Sections 3.1-3.3 are textbook-correct GLRT arguments, and the Boole-Frechet bounds in Section 3.5 are valid as stated. The authors provide code and data via a GitHub repository, which supports reproducibility. However, the significance is substantially weakened by the absence of any comparison to existing QBH systems, the very small evaluation (3 queries, 10 songs), and the circularity of the numerical power claims, so the paper's headline claims are not currently supported.","major_comments":[{"comment":"The abstract claims the system is 'better than existing methods ... empirically in terms of accuracy,' but no existing QBH method is implemented, cited, or compared. Section 5 reports only three hummed queries against a 10-song database containing the query songs, with no baseline. Three queries cannot statistically distinguish the method from chance or from a trivial ranking, and the absence of any QBH literature in the references means the superiority claim has no empirical or contextual grounding.","section":"Abstract and Section 5"},{"comment":"The evaluation is not independent of parameter tuning. The hyperparameters (window length, hop size, neighbour count, threshold criterion) are selected on the first song in Section 5.1, and the same settings are then used for the other two queries. The Jingle Bells run in Section 5.3 (Table 3) shows that Energy and Spectral Dissimilarity detectors, with these settings, rank wrong songs first; only Dominant SD succeeds. Thus the reported success is a fitted summary of the tuning song, not a predictive accuracy, and the claim of general empirical superiority is unsupported.","section":"Section 5.1 and Tables 1-3"},{"comment":"The numerical detection and false-positive probabilities are computed using a noise variance estimated from a blank 1-second recording and a SSNR estimated from the same hummed songs that are later claimed to be detected with at least 90% probability (Section 5.4, Figures 24-26). The mathematical bounds are valid given the model, but the specific numbers are plug-in estimates from the test data, so the 'measure of error' is a fitted parameter, not an independent prediction. The i.i.d. Gaussian noise assumption is also not validated on the actual recordings.","section":"Sections 3.5 and 5.4"},{"comment":"The retrieval pipeline assumes the hummed segment is the same verse as the stored database onsets, and that after a linear tempo rescaling the onset sets correspond one-to-one. Section 4.1.2 states the same verse is present and the timing is equal; Section 4.2 relaxes only timing via a linear transform; Section 7 admits the system cannot handle a user who does not know where the first verse ends. Extra or missing onsets (e.g., meends, ornaments) can cause the subset-matching step to mislabel onsets, so the method is not demonstrated for realistic query variation.","section":"Section 4.1.2, 4.2, and 7"},{"comment":"The claimed advantages in speed and memory are not quantified. The paper states that storing only onsets is memory efficient and that the system is faster than pitch-based methods, but no measurements of runtime, storage size, or comparison against any alternative representation are provided. These claims are load-bearing for the abstract's 'better than existing methods in terms of speed and memory.'","section":"Sections 1 and 6"}],"minor_comments":[{"comment":"There are typos such as 'an sound' and 'getaway' (should be 'way out'), and 'chi-sqaured' appears repeatedly in Section 3.5.1 and 3.5.2.","section":"Section 2.1"},{"comment":"The definition of 'beat' as 'the largest unit of time based on which ... all such measured time in beat units will be integers' is imprecise and does not match standard music terminology; this should be clarified or replaced with a standard definition.","section":"Section 2.2"},{"comment":"The threshold criteria (mean or 3rd quartile of the detection function) are introduced as ad hoc choices with no theoretical justification or sensitivity analysis; a brief discussion of their limitations would improve the paper.","section":"Section 3.4"},{"comment":"The reference list contains no QBH system papers (e.g., Ghias et al., McNab et al., or MIREX QBH participants), which is surprising for a paper claiming superiority over existing QBH methods; at minimum, representative prior work should be cited and discussed.","section":"References"},{"comment":"The statement that 'the squared value of SNR is roughly estimated as 5000 for each of the song' is based on T/σ^2 (mean energy over noise variance) which is not the SNR defined as A/σ; although the paper acknowledges this is a crude underestimate, the notation should be corrected to avoid confusion.","section":"Section 5.4"}],"recommendation":"major_revision","confidential_remarks":"This manuscript reads as a well-executed student project with correct textbook derivations, but it is not yet at the level of a publishable QBH paper. The central comparative claims need either to be supported by a real baseline evaluation (ideally on a standard corpus like MIREX QBH) or substantially softened. The same-verse and equal-timing assumptions are severe limitations that should be stated prominently in the abstract. I recommend major revision rather than rejection because the detection-theoretic core is sound and the onset-only idea may be salvageable, but the current empirical section does not support the claimed contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate proof-of-concept for an onset-only QBH pipeline, and the statistical derivations are mostly solid. But the abstract overstates what the experiments show, and the load-bearing claim of being better than existing QBH systems is untested.\n\nWhat is actually new: using only onset times for query-by-humming, skipping pitch extraction entirely, and combining that with subset and correlative matching. The GLRT derivations for the energy detector and the sinusoid detector are textbook-correct and clearly presented. The matching algorithms are simple, sensible, and cheap. The authors also ship R code and data on GitHub, which makes the work reproducible. They are honest about limitations, including the first-verse requirement and the need for better bounds.\n\nThe soft spots are real and they matter. There is no QBH baseline anywhere in the paper or the references, so the claim of being better than existing methods in speed, memory, and accuracy has no comparison to back it. The experimental evaluation is three hummed queries against a ten-song database, and two of the three detectors fail on Jingle Bells — the paper's best result comes from one detector with hyperparameters tuned on the first song. On a database that small, even a fixed ranking can look reasonable by chance. The power analysis also estimates noise variance and SSNR from the same recordings used to claim detection guarantees, so the quantitative error measure is a fitted summary rather than an independent prediction. The Boole-Frechet bounds are valid as stated, but plugging in estimated parameters weakens the force of the numbers.\n\nI think the stress-test concern is accurate. The same-verse assumption and the linear tempo rescaling are stated plainly in Sections 4.1.2 and 4.2, and Section 7 admits the system cannot handle a user who does not know where the first verse ends. That is a genuine scope limitation, not a hidden flaw, but it further narrows the claim of practical superiority.\n\nThe math and the detection functions themselves hold up. The authors know what they are doing with GLRT and with the peak detection tradeoffs. The weakness is in the evaluation and the comparative claim, not in the core machinery.\n\nWho should read this: anyone working on music retrieval or onset detection who wants a compact example of an onset-only retrieval pipeline, and anyone teaching applied statistics who wants a case study in why baseline comparisons matter. It deserves a serious referee, but the referee should require a larger database, at least one real QBH baseline, and a power analysis done on held-out data before the superiority claim can be taken seriously. My own verdict would be conditional — promising idea, unsupported conclusion as written.","headline":"Onset-only QBH is a plausible idea with sound detection theory, but the headline claim of empirical superiority is not supported by the evidence in the paper.","tokens_in":19108,"tokens_out":1484,"would_cite":false,"duration_ms":17780,"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":"A query-by-humming search engine can identify a song from note-onset timing alone, without storing or analyzing pitch.","keywords":["query by humming","onset detection","dominant spectral dissimilarity","correlative matching","subset matching","rhythm-based retrieval","statistical detection theory","music information retrieval"],"falsifier":"The decisive check is to hum a different verse of the same song, or the same verse with an extra note inserted, and see whether the correlation score still puts the correct song first; the paper's own Section 7 concedes the system cannot locate the end of the first verse, so such a test directly probes the one-to-one matching assumption.","tokens_in":18173,"feed_emoji":"🎵","tokens_out":4506,"duration_ms":43775,"temperature":0.7,"pith_summary":"This paper tries to establish that a Query-by-Humming system can be built from onset times alone, without storing or analyzing pitch. It argues that rhythm carries enough information to identify a song, and that such a system is faster and more memory-efficient than pitch-based QBH and, in the reported experiments, at least as accurate. The authors derive three onset detection functions from detection-theoretic hypotheses, pair them with subset matching and linear-rescaling correlative matching, and test on three hummed songs against a ten-song database. The central claim is that precise onsets are the only information the retrieval step needs.","feed_headline":"Onset times alone can power a query-by-humming system","feed_subtitle":"Storing only the note-onset times of a song's first line shrinks the database and still ranks the right song first.","key_machinery":"The load-bearing objects are the three detection functions, the subset-matching subroutine, and the correlative-matching search. The local energy detector sums squared samples in a sliding window and flags peaks; spectral dissimilarity sums positive magnitude changes across frequency bins below 1 kHz; dominant spectral dissimilarity tracks only positive changes in the maximum squared short-time Fourier magnitude. Subset matching labels each detected onset as a true positive or false positive by nearest-neighbour agreement with the database onsets. Correlative matching tries every linear rescaling of the true onsets onto the output onsets, computes Pearson correlation after subset matching, and multiplies by a correction factor $L^2/(mn)$ that penalizes false positives and false negatives. That correction factor is what keeps the ranking honest about detection errors.","core_discovery":"On the paper's own terms, the discovery is that the timing of note attacks, the onsets, is a sufficient statistic for song identity in a humming-retrieval task. Concretely, the authors claim that storing only the onsets of the first line of each database song, and detecting only onsets in the query, lets a correlation-based ranking recover the correct song in all three experiments with the dominant spectral dissimilarity detector, and in most cases with the other two detectors. The paper also gives a statistical analogy: local energy detection is a generalized likelihood-ratio test for white Gaussian noise against an unknown signal; spectral dissimilarity responds to positive changes across frequency bins; dominant spectral dissimilarity applies a periodogram-maximum test to the dominant frequency. This grounds the detection functions in detection theory rather than pure heuristics.","pith_inferences":["If onset-only matching survives larger and more varied databases, QBH databases could be built automatically from scores or recordings without hiring singers, and queries could be matched even when the user is off-key.","The linear-rescaling assumption could be relaxed by dynamic time warping, which would absorb local tempo drift and ornamented notes instead of forcing one global tempo change.","A testable extension would add coarse pitch or melodic contour only as a secondary reranker, preserving the memory advantage while separating songs whose first-line rhythms are nearly identical."],"forward_implications":["A QBH database can store just a handful of onset times per song rather than audio or MIDI, so scaling to large catalogs becomes mostly a ranking-speed problem.","Because matching is rhythmic rather than melodic, an off-key or imperfect-pitch hum can still retrieve the song as long as note timing is roughly preserved.","Tuning peak detection to avoid false negatives matters more than avoiding false positives, since a missed true onset cannot be recovered at search time.","The dominant spectral dissimilarity detector ranked the correct song first in all three reported experiments; the other two detectors ranked it first in two of three."],"supporting_citations":[{"why":"Supplies the generalized likelihood-ratio detection theory used to derive the energy and dominant spectral dissimilarity detectors.","marker":"[1]"},{"why":"Provides the standard onset detection framework of detection function plus peak picking that the paper adapts.","marker":"[2]"},{"why":"Offers background methods for detecting impulsive events in speech and audio that motivate the onset search.","marker":"[3]"},{"why":"Supplies a pitched-instrument onset detection reference that the paper positions its onset-only approach against.","marker":"[4]"}],"fun_headline_variants":["Just note onsets: a lighter query-by-humming that still ranks","Onset times alone: a leaner, faster query-by-humming","Humming search gets leaner: only note-attack times matter","Onset-only fingerprints: fast, small, accurate humming search","Just the start: onsets alone nail song ID in humming search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The retrieval pipeline depends on the hummed query containing the same verse as the stored database entry and, after a uniform tempo change, on every query onset having one counterpart in the database.","fun_headline_variants_meta":{"raw":{"variants":["Just note onsets: a lighter query-by-humming that still ranks","Onset times alone: a leaner, faster query-by-humming","Humming search gets leaner: only note-attack times matter","Onset-only fingerprints: fast, small, accurate humming search","Just the start: onsets alone nail song ID in humming search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000761,"raw_usage":{"total_tokens":3302,"prompt_tokens":790,"completion_tokens":2512,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":406,"completion_tokens_details":{"reasoning_tokens":2418}},"tokens_in":406,"tokens_out":2512,"duration_ms":16407,"temperature":1.0,"reasoning_tokens":2418,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:49:43.263957+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The decisive check is to hum a different verse of the same song, or the same verse with an extra note inserted, and see whether the correlation score still puts the correct song first; the paper's own Section 7 concedes the system cannot locate the end of the first verse, so such a test directly probes the one-to-one matching assumption.","supporting_citations":[{"cited_title":"Kay, Fundamentals of Statistical Signal Processing, Volume II: Detection Theory","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized likelihood-ratio detection theory used to derive the energy and dominant spectral dissimilarity detectors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the standard onset detection framework of detection function plus peak picking that the paper adapts."},{"cited_title":"Kauppinen, Methods for detecting impulsive noise in speech and audio signals , 2002 14th International Con- ference on Digital Signal Processing Proceedings","cited_arxiv_id":null,"evidence_quote":"Offers background methods for detecting impulsive events in speech and audio that motivate the onset search."},{"cited_title":"Holzapfel, Y","cited_arxiv_id":null,"evidence_quote":"Supplies a pitched-instrument onset detection reference that the paper positions its onset-only approach against."}],"review_version":1}