{"id":"a81349d5-558b-4ea2-947a-f1ac84df5e96","arxiv_id":"1908.06969","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Bayesian piece-specific Markov models that infer sparse repeated rhythms from performed MIDI data improve rhythm transcription, with the second-order metrical model performing best.","lead":"This paper builds Bayesian statistical models of musical rhythm that learn each song's repeated patterns from a performance and use them to transcribe the rhythm into notation. The best model lowers transcription errors on vocal melodies, showing that capturing repetition helps automatic music transcription.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bayesian gains may be due to test-time adaptation, not repetition: the comparison lacks a non-sparse adapted control.","rationale":"The reader's verdict is CONDITIONAL, and I agree the paper should not be rejected. My primary concern is not the simplified 2/4 representation (Sec. 2.1), which the authors explicitly acknowledge (Sec. 5.5). The load-bearing issue is internal to the evaluation: Sec. 4.2 fixes non-Bayesian parameters but estimates Bayesian parameters from the test performance, so the comparison in Sec. 5.3 conflates two variables. The paper's own Sec. 2.3 entropy analysis shows repetition-induced sparsity exists in the corpus, but it does not show that the sparse prior is the causal driver of the observed error-rate reduction. The concentration-parameter study in Sec. 5.4 stays inside the Bayesian family; it cannot separate adaptation from sparsity. The proposed EM control is a minimal addition that would settle the attribution. I therefore keep the verdict CONDITIONAL, and I partially agree with the reader because their rationale mentions a missing external baseline even though their stated weakest assumption is the score representation.","tokens_in":28671,"tokens_out":11566,"duration_ms":124723,"concrete_test":"Using the released code, add a piece-specific non-Bayesian baseline for MetMM2: initialize with the trained generic MetMM2, alternate Viterbi decoding of the test MIDI and maximum-likelihood re-estimation of the transition probabilities from the decoded metrical positions (with the same additive smoothing 0.1 as Sec. 5.1) for a few iterations, then report the final Viterbi error rate on the real data of Sec. 5.3. If this adapted non-sparse baseline's error rate is within one standard deviation of MetMM2B's 2.08%, then the improvement attributed to the sparse repetition prior is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 5.3 compares Bayesian models (e.g. MetMM2B) against non-Bayesian models whose parameters are 'pretrained or preset and fixed during the transcription step' (Sec. 4.2). For the Bayesian models, the piece-specific parameters are instead estimated from the performance being transcribed by Gibbs sampling (Sec. 4.2; Appendix A.2). The comparison therefore varies two things at once: whether the score model adapts to the test performance, and whether a sparse Dirichlet-process prior is used. The conclusion that 'Bayesian score models capturing repetitions' are effective (Sec. 6) requires the gain to come from the repetition-inducing sparsity, but no baseline isolates this. There is no non-Bayesian model adapted to the test piece (e.g. by EM), and Sec. 5.4 only varies the concentration parameter within the Bayesian, adapted family; it does not test a flat or non-sparse adapted model. The entropy analysis in Sec. 2.3 establishes that real piece-specific distributions are sparse, but that is a descriptive property of the corpus, not evidence that the sparse prior causes the transcription improvement. If a non-sparse adapted model performs as well as MetMM2B, the headline result is about adaptation or regularization rather than repetitions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops Bayesian extensions of three families of Markov score models (note-value, metrical, and note-pattern MMs) for rhythm transcription from MIDI. The Bayesian formulation uses Dirichlet-process priors on piece-specific transition/unigram probabilities to induce sparsity, thereby capturing repetitive structure, and optionally models approximate repetitions through onset-shift and note-division processes. The authors derive Gibbs-sampling inference and Viterbi transcription algorithms, and evaluate the models on vocal-melody data from RWC, Beatles, and J-pop corpora using both synthetic and real performed MIDI. The main reported findings are that Bayesian piece-specific models improve transcription accuracy over non-Bayesian generic models for most model types, that the second-order Bayesian metrical Markov model (MetMM2B) is the most accurate, and that modification processes give only small accuracy gains at large computational cost.","tokens_in":28957,"tokens_out":4301,"duration_ms":47908,"significance":"If the central attribution claim holds, this is a useful contribution: it shows that a performance-adaptive, repetition-sensitive score prior can improve rhythm quantization, and it provides a systematic comparison of model families that practitioners could use to choose a tractable model. The paper's strengths include complete formal derivations for the generative models and inference algorithms, an entropy-based analysis connecting repetitive structure to sparsity, and publicly released data and source code. The experimental study is careful in distinguishing synthetic and real data and in reporting standard deviations. However, the headline comparison conflates test-time adaptation with the sparsity/repetition mechanism, and the real-data ranking rests on differences that are partly within statistical fluctuation and on hyperparameters that are partly tuned on the test data. These issues are fixable with additional controls and statistical reporting, but they currently limit the strength of the central claim.","major_comments":[{"comment":"The main comparison is confounded. For non-Bayesian models, the parameters are 'pretrained or preset and are fixed during the transcription step' (Sec. 4.2), whereas for Bayesian models the piece-specific parameters are estimated from the input performance by Gibbs sampling (Sec. 4.2 and Appendix A.2). The comparison therefore varies two factors at once: test-time adaptation of the score model and the sparsity-inducing Dirichlet-process prior that is claimed to capture repetitions. The entropy analysis in Sec. 2.3 shows that real piece-specific distributions are sparse, but that is a descriptive property of the corpus, not evidence that the sparse prior causes the transcription improvement. To support the claim that 'Bayesian score models capturing repetitions' are effective (Sec. 6), the paper needs a non-sparse adapted control, e.g., the same Bayesian inference with a large concentration parameter, a flat or weak Dirichlet prior, or an adapted non-Bayesian model trained by EM. Without such a baseline, the reported gains could be attributed to adaptation or regularization rather than to the repetition mechanism.","section":"Secs. 4.2, 5.3, Fig. 7"},{"comment":"Several evaluation choices make the quantitative claims weaker than the text suggests. The real-data value sigma_t = 0.035 sec is 'roughly optimized' on the real test data (Sec. 5.3 and Fig. 9(c)), and all concentration parameters are fixed to 10 even though Fig. 9(a) shows model-dependent optima; the manuscript states that the differences from the fixed value are within one standard deviation, but the main result should be shown to be stable under the same variation for the other model families. More importantly, no significance tests are reported, and Sec. 5.3 itself concedes that many real-data differences are within statistical fluctuation and that error rates saturate around 2%. Statements such as 'most Bayesian models again significantly outperformed' (Sec. 5.3) and the selection of MetMM2B as 'the best' require pairwise confidence intervals or significance tests, especially because the best model is chosen on the same test set used for hyperparameter tuning.","section":"Secs. 5.1, 5.3, 5.4, Fig. 9"},{"comment":"The scope of the empirical claim should be stated more carefully. The evaluation restricts pieces to 2/4 time on a 16th-note grid, discards rests, onsets finer than a 16th note, segments without onsets, and durations longer than a half note, and assumes a constant, known tempo with Gaussian onset deviations (Secs. 2.1 and 4.1). These restrictions are acknowledged in Sec. 5.5 as a limitation for practical applications, but the abstract's phrase 'universal efficacy of the proposed approach' goes beyond the evidence, which covers monophonic vocal melodies under a simplified score representation. I recommend softening the abstract and conclusion to state that the demonstrated gains hold for the tested representation and corpus, with extension to more general settings left as future work.","section":"Secs. 2.1, 5.5, Abstract"}],"minor_comments":[{"comment":"The beam width for PatMM1DB and PatMM1SDB is reported as W = 200 in Sec. 5.3, but Table 4's caption states W = 1000 for the same models; please reconcile this discrepancy.","section":"Sec. 5.3 vs. Table 4"},{"comment":"There is a typo in the definition of the latent variables: 'zn = (kn.in,sn)' should read 'zn = (kn,in,sn)'.","section":"Supplemental Material, Sec. 4.2.2"},{"comment":"The figure compares the 'theoretically optimal' concentration parameters from entropy matching with the empirically optimal values, but no explanation is given for why the entropy-matched values are not expected to coincide with transcription-optimal values; a sentence describing this mismatch would help the reader interpret the plot.","section":"Sec. 5.4, Fig. 9(a)"},{"comment":"The modification hyperparameters are preset to ¯ξ0 = ¯ζ0 = 0.9 with no sensitivity analysis; since modification models are a central modeling component, reporting the effect of varying these values would strengthen the paper, even if only for one model family.","section":"Sec. 4.2"},{"comment":"The example in Fig. 8 is informative, but the selection criterion ('only bars with repeated rhythms') should be stated explicitly, and the transcription error markers should be described in the caption for clarity.","section":"Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically competent, with clear derivations and a useful public release of data and code. The main concern is that the central claim's causal attribution to 'repetitions' is not identifiable from the current experiments, because test-time adaptation is confounded with the sparse prior. This is fixable with an additional control experiment, which I would expect to be feasible within the paper's scope. The self-citation to [18] is appropriate, as the present work is a systematic extension of that earlier model. I see no novelty-disclosure concern; the main question is whether the empirical design can be tightened enough to support the headline claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, reproducible extension of the authors' own Dirichlet-process score-model idea, and people working on rhythm transcription will get real value from the systematic comparison. But don't take the abstract's 'universal efficacy' literally. The experiments vary adaptation to the test piece and sparsity at the same time, so the paper does not actually isolate repetitions as the cause of the gains.\n\nWhat's new: Bayesian formulations of note-value and metrical Markov models, onset shifts generalized to any note, and a broad model comparison with a clear practical recommendation (MetMM2B). The Gibbs sampling derivations are standard but complete, code and data are available, and the held-out evaluation is refreshingly honest—the authors manually inspect the remaining errors and attribute most to performance mistakes. The entropy analysis convincingly shows real piece-specific distributions are sparser than finite-sample chance, which supports the descriptive premise.\n\nSoft spots. The main one is the stress-test concern, and I think it lands. The non-Bayesian baselines have fixed generic parameters; the Bayesian models estimate piece-specific parameters from the test input. So 'Bayesian beats non-Bayesian' conflates adaptation with sparsity. A non-sparse adapted control—for example, EM with a flat prior or a finite mixture fitted to the test piece—is needed to support the claim that the improvement comes from capturing repetitions. The concentration sweep in Sec. 5.4 stays inside the sparse family and doesn't fix this. Without that control, the paper supports 'piece-specific adaptation helps,' not specifically 'sparsity from repetitions helps.'\n\nThe other issues are minor to moderate. sigma_t is roughly tuned on the real test data, all concentration parameters are fixed to 10 even though Sec. 5.4 shows model-dependent optima, and there are no significance tests. On the real data, the differences among the top models look close to statistical noise. The task setup is also simplified: 2/4 meter, 16th-note grid, no rests, constant known tempo, monophonic melodies. That is acceptable for a controlled comparison, but the practical transfer claim should be softened. The modification models add little accuracy and large computational cost—the paper says this clearly, which I credit.\n\nBottom line: it deserves a serious peer review, with two requests: a non-sparse adapted baseline, and proper statistical reporting. As it stands, this is an incremental but genuinely useful contribution, best read as 'adaptive score models are worth using; the second-order Bayesian metrical model is the efficient choice.' I would bring it to our reading group and would cite it for the reproducible comparison of model families.","headline":"Useful, reproducible extension of the authors' DP score-model line, but the headline claim about repetitions is confounded with test-time adaptation and needs a non-sparse adapted control.","tokens_in":29415,"tokens_out":2546,"would_cite":true,"duration_ms":28973,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Rhythm transcription improves when the score model is inferred per piece from the performance itself, and the best tested model is the second-order Bayesian metrical Markov model.","keywords":["music transcription","rhythm transcription","Bayesian score models","Dirichlet process","Markov models","repetitive structure","MIDI performance","metrical Markov model"],"falsifier":"Run the best model, MetMM2B, on a test corpus that includes triplet rhythms, rests, tempo changes, or rubato; if the repetition-driven Bayesian model no longer beats a generic non-Bayesian baseline, or if the error rate rises well above the roughly 2% reported for simplified data, the piece-specific repetition mechanism is not doing the work in realistic settings.","tokens_in":28509,"feed_emoji":"🎵","tokens_out":7734,"duration_ms":66857,"temperature":0.7,"pith_summary":"This paper claims that rhythm transcription improves when the statistical score model is tailored to the piece being transcribed rather than fixed across all music. The authors formulate Bayesian Markov models whose transition probabilities are drawn from a Dirichlet process with a generic base distribution and a small concentration parameter, so each piece receives a sparse distribution of rhythmic units and repetition emerges naturally. On vocal melodies from popular music, the Bayesian versions outperform their non-Bayesian counterparts for most model types, and the second-order Bayesian metrical Markov model gives the lowest error rate, about 2% on real performances. If the claim holds, the repetitive structure present in a performance is itself a reliable guide for transcription.","feed_headline":"Repetition-aware Bayesian models cut rhythm transcription errors","feed_subtitle":"A Bayesian model uses each piece's repeated rhythms to beat generic baselines on MIDI.","key_machinery":"The load-bearing object is the Dirichlet process prior placed on the parameters of a Markov score model: for each piece, the initial and transition distributions (over note values, metrical positions, or note patterns) are drawn from $\\mathrm{DP}(\\alpha, \\bar{\\pi})$ with a small concentration parameter $\\alpha$, making the per-piece distribution sparse and thereby encoding repetition without an explicit repeated-structure grammar. Inference couples Gibbs sampling, which estimates the piece-specific parameters from the input MIDI, with a Viterbi (or beam-searched Viterbi) step that decodes the most probable note values or metrical positions. Two note-modification operations, onset shifts and note divisions, are added as hidden variables to represent approximate repetitions.","core_discovery":"On its own terms, the paper establishes that a piece-specific score model inferred from the performed MIDI data, rather than a generic score model trained on other pieces, captures the repetitions that dominate real rhythms. The sparseness of these piece-specific distributions is quantitatively visible: their entropies are far lower than those of finite samples from the generic model, and the Dirichlet process with a small concentration parameter reproduces the observed entropy range. The central experimental result is that Bayesian extensions of the note-value, metrical, and note-pattern Markov models significantly reduce transcription error rates on both synthetic and human-performed MIDI, with MetMM2B the most accurate at roughly 2% error on real data. The paper also shows that explicitly modeling approximate repetitions through onset shifts and note divisions yields modest accuracy gains that do not justify their large computational cost for the model types tested.","pith_inferences":["A direct extension would test whether per-piece concentration parameters, learned from the input rather than fixed, push the error rate below the observed ~2% floor on real performances.","The same Dirichlet-process piece-specific mechanism could be applied to pitch or chord sequences, where repetition is also common; if the rhythm results transfer, score models for full polyphonic transcription could be built the same way.","Because the paper's simplification drops rests and triplets, a 'simplified score world' bias may inflate the benefit of repetition; a falsifying test is to add such rhythms and see whether the generic baseline catches up.","One could measure the learned piece-specific distributions directly: if the sparse-transition mechanism is truly active, the inferred concentration parameters for pieces with little repetition should approach the generic case, which would be a checkable signature."],"forward_implications":["Bayesian piece-specific score models lower rhythm transcription error for most model types, often by more than raising the Markov order of a generic model.","MetMM2B gives the best accuracy-computation trade-off and is the recommended practical model; note-pattern models are most accurate only at prohibitive cost.","Explicitly modeling onset shifts and note divisions captures approximate repetitions but yields only small accuracy gains, so simpler Bayesian models without modifications are attractive.","The approach extends to finer beat grids and longer note values for metrical MMs, since their cost grows polynomially in bar length, not exponentially as with note-pattern MMs.","The same piece-specific Bayesian construction can be applied to other musical elements, such as pitches, chords, or polyphonic voices, where repetitions also guide recognition (stated as future work in the paper)."],"supporting_citations":[{"why":"Supplies the prior hierarchical Bayesian score model and the note-modification operations that this work generalizes and extends.","marker":"[18]"},{"why":"Provides the Dirichlet process mechanism used to generate sparse piece-specific distributions from a generic base distribution.","marker":"[14]"},{"why":"Supplies the topic-model analogy, where each document gets its own word distribution, that motivates piece-specific score models.","marker":"[5]"},{"why":"Defines the note-value Markov model, one of the three base score model classes given a Bayesian extension.","marker":"[31]"},{"why":"Define the metrical Markov model, the base class of the best-performing Bayesian model MetMM2B.","marker":"[10, 23]"},{"why":"Provides the Gibbs sampling procedure used for posterior inference of the piece-specific model parameters.","marker":"[4]"},{"why":"Supplies the Viterbi algorithm used to decode the most probable score from the trained HMM.","marker":"[21]"},{"why":"Supplies the RWC popular music dataset used as part of the training and test material for the experiments.","marker":"[9]"}],"fun_headline_variants":["Bayesian models use each piece's repetitions to improve rhythm transcription","Piece-specific Bayesian score models capture repetitions for better rhythm transcription","Repetition-capturing Bayesian models cut rhythm transcription errors","Piece-specific Bayesian models reduce rhythm transcription errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes a simplified score world—2/4 time on a 16th-note grid, no rests, no triplets, no segments without onsets, and a constant known tempo with Gaussian onset deviations—and if the real rhythmic phenomena excluded here matter, the measured improvements may not transfer to practical transcription.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian models use each piece's repetitions to improve rhythm transcription","Piece-specific Bayesian score models capture repetitions for better rhythm transcription","Repetition-capturing Bayesian models cut rhythm transcription errors","Piece-specific Bayesian models reduce rhythm transcription errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3342,"prompt_tokens":907,"completion_tokens":2435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":2369}},"tokens_in":523,"tokens_out":2435,"duration_ms":17052,"temperature":1.0,"reasoning_tokens":2369,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:44:59.214066+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the best model, MetMM2B, on a test corpus that includes triplet rhythms, rests, tempo changes, or rubato; if the repetition-driven Bayesian model no longer beats a generic non-Bayesian baseline, or if the error rate rises well above the roughly 2% reported for simplified data, the piece-specific repetition mechanism is not doing the work in realistic settings.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the prior hierarchical Bayesian score model and the note-modification operations that this work generalizes and extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Dirichlet process mechanism used to generate sparse piece-specific distributions from a generic base distribution."},{"cited_title":"M., Ng, A","cited_arxiv_id":null,"evidence_quote":"Supplies the topic-model analogy, where each document gets its own word distribution, that motivates piece-specific score models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the note-value Markov model, one of the three base score model classes given a Bayesian extension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Gibbs sampling procedure used for posterior inference of the piece-specific model parameters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RWC popular music dataset used as part of the training and test material for the experiments."}],"review_version":1}