Pith. sign in

REVIEW 4 major objections 4 minor 32 references

A new violin transcription system, VioPTT, jointly outputs pitch, onset, offset, and playing technique, and its technique head—trained only on synthetic audio—achieves 77% macro accuracy on real recordings.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A cascade model transcribes violin pitch, onset, offset, and playing technique, trained on 76 hours of synthetic VST-rendered audio.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection First violin technique-aware transcription system with a useful synthetic dataset, but the joint model is never tested end-to-end and the velocity ablation is internally inconsistent. the 4 major comments →

arxiv 2509.23759 v3 pith:ZSIR6TD2 submitted 2025-09-28 cs.SD cs.LG

VioPTT: Violin Technique-Aware Transcription from Synthetic Data Augmentation

Ting-Kang Wang , Yueh-Po Peng , Li Su , Vincent K.M. Cheung This is my paper
classification cs.SD cs.LG
keywords automatic music transcriptionviolin playing techniquearticulation classificationsynthetic data augmentationmusic information retrievaltechnique-aware transcriptionsolo violinMOSA-VPT
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to extend automatic music transcription for the violin so that a single model outputs not only which notes are played and when, but also how each note is played—détaché, flageolet, spiccato, or pizzicato. The blocker is annotation cost: technique labels require expert judgment and do not exist at scale. The authors sidestep it by rendering 76 hours of technique-labeled audio from MIDI scores using a virtual violin, training the technique recognizer on that synthetic corpus alone, and showing it transfers to real recordings. On real test data they report state-of-the-art pitch and timing transcription on two benchmarks and 77.22% macro accuracy for technique classification, alongside the first unified joint transcription-and-technique output. The payoff, if the transfer claim holds up, is that expressive performance attributes can be added to transcription without waiting for expensive human labels.

Core claim

The paper's central claim is that a lightweight, cascade model—a frame-level transcription module followed by a note-level articulation module—can jointly transcribe pitch, onset, offset, and playing technique for violin, and that the technique head can be trained entirely on synthetic audio yet classify real performances at high accuracy. The synthetic corpus MOSA-VPT is produced by rendering MIDI through a professional virtual violin with key-switch and controller control for four techniques; no expert labels are involved. The authors report that this synthetic-only training yields 77.22% macro accuracy on the RWC real-instrument corpus, and that the full model exceeds prior state-of-the-a

What carries the argument

The load-bearing mechanism is the fusion of two feature streams in the articulation module: a 128-dimensional acoustic embedding from a convolutional encoder over log-mel spectra, concatenated with a 128-dimensional embedding of the transcription module's frame-level predictions (onset, offset, frame, velocity). This fused representation is classified into four techniques plus 'no technique'. The data side is MOSA-VPT, a 76-hour synthetic corpus rendered with a VST violin whose key switches and continuous controllers produce technique-labeled note events; rendering is deliberately stripped of room processing and mixed to mono at 16 kHz to reduce domain bias.

Load-bearing premise

The technique classifier transfers from synthetic VST renderings to real violin recordings, even though the paper's own embedding visualization shows synthetic and real clusters diverging.

What would settle it

Evaluate the articulation module trained only on MOSA-VPT on a new set of real solo-violin recordings with expert labels for the four techniques; if macro accuracy drops far below the reported 77% or confusion concentrates between détaché and spiccato, the synthetic-transfer premise fails. A simpler probe: the ablation shows removing velocity kills pizzicato accuracy (to 0.16%), so real pizzicato recordings with intentionally flat dynamics should be tested; if they are still classified correctly, the velocity dependence is not causal, and if not, the model is relying on a synthetic artifact.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If correct, a violin transcription system can output expressive technique labels at note level, not just pitch and timing, from a single forward pass.
  • Expert technique annotation is not required for training; synthetic rendering can substitute at scale for bowed-string techniques.
  • The ablation results indicate that different techniques rely on different cues—pizzicato on velocity, flageolet on harmonic/frame content, bow-stroke techniques on onset/offset timing—a finding that should guide future technique-aware models.
  • The synthesis pipeline itself is instrument-agnostic; any VST with articulation control could produce technique-labeled corpora for other instruments.
  • The model matches or exceeds state-of-the-art violin transcription on URMP and Bach10 despite using about 30% less training data than the MUSC baseline, attributing the gain to data augmentation and well-aligned supervision.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The UMAP divergence reported in the paper (détaché closer to flageolet, spiccato closer to pizzicato) suggests that the synthetic-to-real transfer is partial; a reader should expect technique confusion to be higher on real performances than on synthetic ones, especially between détaché and spiccato.
  • A natural next test is to apply the same synthetic-only recipe to a post-tonal or contemporary violin technique (e.g., sul ponticello, sul tasto, col legno) to see whether the VST's rendering preserves these cues; if it does not, the method's coverage will be limited to the four rendered techniques.
  • The technique head's near-zero pizzicato accuracy when velocity is ablated suggests that the model may be using a velocity shortcut rather than timbral content for pizzicato; real pizzicato recordings with varied dynamics would test whether that cue generalizes.
  • Because the transcription and articulation modules are trained separately, end-to-end joint training could shift representations and improve or worsen technique accuracy; this is an untested alternative within the paper's own architecture.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes VioPTT, a lightweight cascade for violin transcription that predicts pitch, onset, offset, and playing technique (détaché, flageolet, spiccato, pizzicato). The transcription module follows Kong et al.'s high-resolution piano architecture, trained on the MOSA dataset with pitch/timing augmentation; the articulation module is trained on MOSA-VPT, a newly synthesized 76-hour dataset rendered with a commercial violin VST. Transcription is evaluated on URMP and Bach10 against MUSC, and technique classification on isolated RWC notes with feature ablations. The paper claims state-of-the-art transcription and successful synthetic-to-real transfer for technique prediction.

Significance. If the claims hold, VioPTT would be a useful step toward expressive AMT, and MOSA-VPT would be a valuable resource that avoids costly expert technique annotations. Strengths include evaluation on external public datasets (URMP, Bach10, RWC), comparison to an existing violin transcription baseline (MUSC), a thoughtful ablation of transcription features for technique classification, and a data-synthesis pipeline that is, in principle, extendable to other instruments. However, the central claim of joint technique-aware transcription in real performances is not actually tested end-to-end, and one ablation result conflicts with the stated implementation. These issues need to be resolved before the main conclusions can be accepted.

major comments (4)
  1. [§4.2, §5.2, §6] The central claim that VioPTT 'transcribes playing technique in real-life performances' (Section 6) is not supported by the experiments. The articulation module is trained and evaluated on 2-second isolated single notes (Section 4.2), and the RWC subset is explicitly 'separated into single notes' (Section 4.1). Table 2 therefore measures a single-note classifier, not the output of the full cascade on continuous audio. To substantiate the joint-transcription claim, the authors should evaluate the complete model on continuous recordings, e.g., by running the transcription module on RWC scale performances or another annotated real-violin recording and reporting note-level technique accuracy for the detected notes. Without this, the paper's headline contribution—joint pitch/onset/offset/technique transcription—is untested. Additionally, no comparison is made to existing technique classifiers
  2. [§4.2 vs. Table 2] There is an internal inconsistency that undermines the velocity ablation. Section 4.2 states: 'As with most transcription models, we omit velocity transcription and set it to a constant value.' If the velocity feature is constant, removing it ('Velocity excluded' in Table 2) should have no effect. Yet the table shows Pizzicato accuracy dropping from 88.80% to 0.16% when velocity is excluded. This is impossible under the stated implementation. Either the velocity input to the articulation module is not constant, or the ablation removes something else. The interpretation in Section 5.2 that 'the model may have exploited velocity as a defining characteristic for this technique' is not valid without resolving this contradiction.
  3. [Table 1, §5.1, Abstract] The transcription results are reported as point estimates with no variance or repeated-run statistics. Several differences are small (e.g., URMP F1: MUSC 84.6 vs. Ours w/ aug 84.5), so the claims of 'exceeding the state-of-the-art' and 'matching' MUSC are not statistically grounded. Moreover, the abstract and conclusion say 'exceeding the state-of-the-art', but on URMP the proposed model does not numerically exceed MUSC in F1 (84.5 vs. 84.6); only Bach10 shows a clear improvement. The authors should either provide error bars/statistical tests or temper the claim to 'comparable or better' where appropriate.
  4. [Fig. 3, §5.2] The synthetic-to-real transfer evidence is only partially quantitative. Figure 3 shows that real embeddings diverge from synthetic ones—détaché clusters near flageolet and spiccato near pizzicato—which weakens the assumption that the synthetic renderings preserve the acoustic cues needed for transfer. The paper's interpretation that 'all four classes were largely disentangled' is qualitative; no quantitative measure of domain shift or cluster separability is provided. Given that the entire technique module is trained on synthetic data, a quantitative domain-gap analysis (e.g., per-class centroid distances, or a supervised domain-discrimination test) would strengthen the central transfer claim.
minor comments (4)
  1. [Table 2 caption] 'perfromance' should be 'performance'.
  2. [§3.1] 'transciption' should be 'transcription'.
  3. [§6] 'more wider variety' is ungrammatical; should be 'a wider variety'.
  4. [References [18], [23]] The references for DAWDreamer and the Bach10 dataset lack full bibliographic information (venue, year, or publisher); please complete them.

Circularity Check

0 steps flagged

No significant circularity: the reported predictions are benchmarked on external public datasets and the synthetic training data is independent of the test sets.

full rationale

The paper's derivation chain is self-contained rather than circular. The transcription module is trained on MOSA real solo-violin recordings and evaluated on URMP and Bach10, while the baseline MUSC and all test labels are external to the training pipeline. The articulation module is trained on the newly synthesized MOSA-VPT corpus and evaluated on RWC real single-note recordings; no parameter is fitted to RWC or to the reported macro accuracy, so the technique results are a genuine transfer test rather than a fitted-input prediction. The self-citations (Su et al. 2014 and Huang et al. 2023, both involving co-author Li Su) appear only in related-work context and are not load-bearing assumptions or imported uniqueness theorems. The claim of being the first joint violin transcription and technique system is a novelty statement, not a derivation. Two validity concerns exist but are not circularity: Section 4.2 says velocity is set to a constant, while Table 2's 'Velocity excluded' row reports a large effect, an internal inconsistency; and the technique module is evaluated on isolated single notes rather than on full continuous performances, so the end-to-end technique-transcription claim is not fully demonstrated. These affect experimental support, but neither makes any reported result reduce by construction to its own input.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 1 invented entities

The central claims rest on a synthetic-data training recipe whose hyperparameters are hand-picked but standard, and on the domain assumption that VST renderings transfer to real acoustics. No new theoretical entities are introduced; the dataset is a new resource but not yet publicly available.

free parameters (1)
  • Pitch/timing augmentation hyperparameters = pitch shift ±0.1 semitones, gain +5 dB, band-pass 32-4096 Hz, room size 0.35
    Hand-chosen in Section 3.2.1; not fitted, but part of the training recipe. The central claim does not rest on exact values, but they are additional tunable constants.
axioms (4)
  • domain assumption RWC playing-style labels map onto the four technique classes (détaché, flageolet, spiccato, pizzicato) used in MOSA-VPT.
    Section 5.2 evaluates on RWC using these classes; no verification of label equivalence is provided.
  • domain assumption Synchron Solo Violin I VST renderings with key switches and CCs are acoustically representative of real violin techniques.
    Section 3.2.2; Fig. 3 shows synthetic and real embeddings diverge, casting doubt on this premise.
  • domain assumption MOSA note annotations are accurate enough for training the transcription module.
    Section 4.1 uses MOSA as core training data; annotation accuracy is not independently audited.
  • domain assumption Standard note-matching criteria (50 cents, 50 ms, 20% offset) and mir_eval are correctly applied.
    Section 4.2 defines the metrics; no cross-check or error analysis is provided.
invented entities (1)
  • MOSA-VPT synthetic dataset no independent evidence
    purpose: Provides note- and technique-aligned training data without expert annotation (76 hours, four techniques)
    Release is pending ('link to dataset will be made available upon acceptance', Section 4.1); the paper's RWC results are internal, not independent confirmation.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of VioPTT: Violin Technique-Aware Transcription from Synthetic Data Augmentation." pith.science (2026). https://pith.science/paper/ZSIR6TD2

@misc{pith2026250923759,
  author       = {Pith},
  title        = {Pith review of: VioPTT: Violin Technique-Aware Transcription from Synthetic Data Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZSIR6TD2}},
  note         = {Machine review of arXiv:2509.23759}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

While automatic music transcription is well-established in music information retrieval, most models are limited to transcribing pitch and timing information from audio, and thus omit crucial expressive and instrument-specific nuances. One example is playing technique on the violin, which affords its distinct palette of timbres for maximal emotional impact. Here, we propose VioPTT (Violin Playing Technique-aware Transcription), a lightweight cascade model that directly transcribes violin playing technique in addition to pitch onset and offset. Furthermore, we release MOSA-VPT, a novel, high-quality synthetic violin playing technique dataset to circumvent the need for manually labeled annotations. Leveraging this dataset, our model demonstrated strong generalization to real-world note-level violin technique recordings in addition to achieving state-of-the-art transcription performance. To our knowledge, VioPTT is the first to jointly combine violin transcription and playing technique prediction within a unified framework.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 5 linked inside Pith

  1. [1]

    As one of the most studied tasks in music in- formation retrieval (MIR), AMT has achieved substantial progress with the advent of deep learning

    INTRODUCTION The goal of automatic music transcription (AMT) is to convert an audio recording into a symbolic representation containing pitch and timing information. As one of the most studied tasks in music in- formation retrieval (MIR), AMT has achieved substantial progress with the advent of deep learning. In particular, instrument-specific models have...

  2. [2]

    RELA TED WORK Despite the recent emergence of instrument-agnostic transcription models (e.g., [8, 9]), piano transcription remains the most widely studied domain within AMT. The high-resolution model by Kong et al.[1] achieved state-of-the-art performance on the MAESTRO arXiv:2509.23759v2 [cs.SD] 30 Sep 2025 dataset [10] by modeling onset and offset times...

    Pith/arXiv arXiv 2025

  3. [3]

    Playing Technique-Aware Transcription Model As illustrated in Fig

    METHOD 3.1. Playing Technique-Aware Transcription Model As illustrated in Fig. 1, our model comprises two components: a transcription modulethat predicts note information at the frame level, and anarticulation modulethat assigns a technique label to each transcribed note. The input is a log mel-spectrogram tensor with different STFT window lengths (length...

  4. [4]

    First, we examined the impact of transfer learning and data augmentation (see Section 3.2.1) on transcribing pitch and timing information

    EXPERIMENTS We conducted two experiments to evaluate transcription perfor- mance of our model. First, we examined the impact of transfer learning and data augmentation (see Section 3.2.1) on transcribing pitch and timing information. To test whether knowledge from a related domain could benefit violin transcription, we initialized the transcription module...

  5. [5]

    FT” denotes models fine-tuned from a piano-pretrained checkpoint; “aug

    RESULTS AND DISCUSSION 5.1. Pitch and Timing Transcription Table 1 reports pitch and timing transcription performance with and without piano pretraining, as well as with and without data augmen- tation. On URMP, our model achieved the highest Recall (83.6) and F1no (93.1), while also matching the state-of-the-art MUSC [7] in Precision and note-level F1 wh...

  6. [6]

    CONCLUSION AND FUTURE WORK While pitch and timing information remain fundamental in AMT, transcribing expressive nuances from different playing techniques is a crucial next step towards fine-grained music performance modeling. Here, we presented VioPTT, a light-weight end-to-end technique-aware transcription model for violin, and release MOSA- VPT, a high...

  7. [7]

    and the National Science and Technology Coun- cil, Taiwan under grant MOST 110-2221-E-001-010-MY3

    ACKNOWLEDGMENTS This work was supported by funding from Sony Computer Science Laboratories, Inc. and the National Science and Technology Coun- cil, Taiwan under grant MOST 110-2221-E-001-010-MY3

  8. [8]

    High-resolution piano transcription with ped- als by regressing onset and offset times,

    Qiuqiang Kong, Bochen Li, Xuchen Song, Yuan Wan, and Yuxuan Wang, “High-resolution piano transcription with ped- als by regressing onset and offset times,”IEEE/ACM Transac- tions on Audio, Speech, and Language Processing, vol. 29, pp. 3707–3717, 2021

    2021

  9. [9]

    Automatic piano transcription with hierarchical frequency-time trans- former,

    Keisuke Toyama, Taketo Akama, Yukara Ikemiya, Yuhta Takida, Wei-Hsiang Liao, and Yuki Mitsufuji, “Automatic piano transcription with hierarchical frequency-time trans- former,”arXiv preprint arXiv:2307.04305, 2023

    Pith/arXiv arXiv 2023

  10. [10]

    Onsets and frames: Dual-objective piano transcription,

    Curtis Hawthorne, Erich Elsen, Jialin Song, Adam Roberts, Ian Simon, Colin Raffel, Jesse Engel, Sageev Oore, and Douglas Eck, “Onsets and frames: Dual-objective piano transcription,” arXiv preprint arXiv:1710.11153, 2017

    Pith/arXiv arXiv 2017

  11. [11]

    Sequence-to-sequence piano transcription with transformers,

    Curtis Hawthorne, Ian Simon, Rigel Swavely, Ethan Manilow, and Jesse Engel, “Sequence-to-sequence piano transcription with transformers,” inProceedings of the International Society for Music Information Retrieval Conference (ISMIR), 2021

    2021

  12. [12]

    Note and playing technique transcription of electric guitar solos in real- world music performance,

    Tung-Sheng Huang, Ping-Chung Yu, and Li Su, “Note and playing technique transcription of electric guitar solos in real- world music performance,” inICASSP 2023-2023 IEEE Inter- national Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  13. [13]

    High resolution guitar transcription via domain adaptation,

    Xavier Riley, Drew Edwards, and Simon Dixon, “High resolution guitar transcription via domain adaptation,” in ICASSP 2024-2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 1051–1055

    2024

  14. [14]

    High-resolution violin transcription using weak labels,

    Nazif Can Tamer, Yigitcan ¨Ozer, Meinard M ¨uller, and Xavier Serra, “High-resolution violin transcription using weak labels,” inIsmir 2023 Hybrid Conference, 2023

    2023

  15. [15]

    A lightweight instrument-agnostic model for polyphonic note transcription and multipitch estimation,

    Rachel M. Bittner, Juan Jos ´e Bosch, David Rubinstein, Gabriel Meseguer-Brocal, and Sebastian Ewert, “A lightweight instrument-agnostic model for polyphonic note transcription and multipitch estimation,” inProceedings of the IEEE In- ternational Conference on Acoustics, Speech, and Signal Pro- cessing (ICASSP), Singapore, 2022

    2022

  16. [16]

    Mt3: Multi-task multitrack music transcrip- tion,

    Josh Gardner, Ian Simon, Ethan Manilow, Curtis Hawthorne, and Jesse Engel, “Mt3: Multi-task multitrack music transcrip- tion,”arXiv preprint arXiv:2111.03017, 2021

    Pith/arXiv arXiv 2021

  17. [17]

    Enabling factorized piano music modeling and generation with the MAESTRO dataset,

    Curtis Hawthorne, Andriy Stasyuk, Adam Roberts, Ian Simon, Cheng-Zhi Anna Huang, Sander Dieleman, Erich Elsen, Jesse Engel, and Douglas Eck, “Enabling factorized piano music modeling and generation with the MAESTRO dataset,” inPro- ceedings of the International Conference on Learning Repre- sentations (ICLR), 2019

    2019

  18. [18]

    Skipping the frame-level: Event-based piano transcription with neural semi- crfs,

    Yujia Yan, Frank Cwitkowitz, and Zhiyao Duan, “Skipping the frame-level: Event-based piano transcription with neural semi- crfs,”Advances in Neural Information Processing Systems, vol. 34, pp. 20583–20595, 2021

    2021

  19. [19]

    Scoring time intervals us- ing non-hierarchical transformer for automatic piano transcrip- tion,

    Yujia Yan and Zhiyao Duan, “Scoring time intervals us- ing non-hierarchical transformer for automatic piano transcrip- tion,”arXiv preprint arXiv:2404.09466, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Note- level automatic guitar transcription using attention mecha- nism,

    Sehun Kim, Tomoki Hayashi, and Tomoki Toda, “Note- level automatic guitar transcription using attention mecha- nism,” in2022 30th European Signal Processing Conference (EUSIPCO). IEEE, 2022, pp. 229–233

    2022

  21. [21]

    Statistical modeling of bowing control ap- plied to violin sound synthesis,

    Esteban Maestre, Merlijn Blaauw, Jordi Bonada, Enric Guaus, and Alfonso P´erez, “Statistical modeling of bowing control ap- plied to violin sound synthesis,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 18, no. 4, pp. 855–871, 2010

    2010

  22. [22]

    Sparse modeling of magnitude and phase-derived spectra for playing technique classification,

    Li Su, Hsin-Ming Lin, and Yi-Hsuan Yang, “Sparse modeling of magnitude and phase-derived spectra for playing technique classification,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 22, no. 12, pp. 2122–2132, 2014

    2014

  23. [23]

    Playing technique classification for bowed string instruments from raw audio,

    AB Kruger and JP Jacobs, “Playing technique classification for bowed string instruments from raw audio,”Journal of New Music Research, vol. 49, no. 4, pp. 320–333, 2020

    2020

  24. [24]

    Audio classification of violin bowing techniques: An aid for beginners,

    Hernan S Alar, Ramil O Mamaril, Lex P Villegas, and Jhon Roe D Cabarrubias, “Audio classification of violin bowing techniques: An aid for beginners,”Machine Learning with Applications, vol. 4, pp. 100028, 2021

    2021

  25. [25]

    DawDreamer: Bridging the Gap Between Dig- ital Audio Workstations and Python Interfaces,

    David Braun, “DawDreamer: Bridging the Gap Between Dig- ital Audio Workstations and Python Interfaces,”

  26. [26]

    Mosa: Music mo- tion with semantic annotation dataset for cross-modal music processing,

    Yu-Fen Huang, Nikki Moran, Simon Coleman, Jon Kelly, Shun-Hwa Wei, Po-Yin Chen, Yun-Hsin Huang, Tsung-Ping Chen, Yu-Chia Kuo, Yu-Chi Wei, et al., “Mosa: Music mo- tion with semantic annotation dataset for cross-modal music processing,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 4157–4170, 2024

    2024

  27. [27]

    Creating a multitrack classical music per- formance dataset for multimodal music analysis: Challenges, insights, and applications,

    Bochen Li, Xinzhao Liu, Karthik Dinesh, Zhiyao Duan, and Gaurav Sharma, “Creating a multitrack classical music per- formance dataset for multimodal music analysis: Challenges, insights, and applications,”IEEE Transactions on Multimedia, vol. 21, no. 2, pp. 522–535, 2018

    2018

  28. [28]

    PYIN: A fundamen- tal frequency estimator using probabilistic threshold distribu- tions,

    Matthias Mauch and Simon Dixon, “PYIN: A fundamen- tal frequency estimator using probabilistic threshold distribu- tions,” inProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2014, pp. 659–663

    2014

  29. [29]

    Computer-aided melody note transcription using the tony software: Accuracy and efficiency,

    Matthias Mauch, Chris Cannam, Rachel Bittner, George Fazekas, Justin Salamon, Jiajie Dai, Juan Bello, and Simon Dixon, “Computer-aided melody note transcription using the tony software: Accuracy and efficiency,” inProceedings of the International Conference on Technologies for Music Notation and Representation, 2015

    2015

  30. [30]

    Bach10 dataset,

    A Versatile Polyphonic Music Dataset, “Bach10 dataset,”

  31. [31]

    Rwc music database: Music genre database and musical instrument sound database,

    Masataka Goto, Hiroki Hashiguchi, Takuichi Nishimura, and Ryuichi Oka, “Rwc music database: Music genre database and musical instrument sound database,” 2003

    2003

  32. [32]

    MIR EV AL: A transparent implementation of common mir metrics,

    Colin Raffel, Brian McFee, Eric J Humphrey, Justin Salamon, Oriol Nieto, Dawen Liang, and Daniel PW Ellis, “MIR EV AL: A transparent implementation of common mir metrics,” inPro- ceedings of the International Society for Music Information Re- trieval Conference (ISMIR), 2014

    2014

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.