Source-linked AI summary
Gangqinpu: A Multi-Genre, Multi-Task Piano Dataset with Aligned Audio, MIDI, and Musical Scores
Wei Zeng, Anvay Mathur, Ye Wang
TL;DR
Datasets jointly aligning piano audio, performance MIDI, and scores remain scarce, particularly across genres relevant to music analysis and performance modeling. Gangqinpu constructs a large multi-genre aligned dataset with automatic annotations and leakage-aware splits. Experiments report beat-tracking gains on both Gangqinpu and ASAP, while transcription and rendering gains are primarily in-domain.
Problem
Piano datasets that jointly align audio, performance MIDI, and scores remain scarce, and existing repertoire coverage is often limited in scale and genre diversity.
Source: §1 · p. 1Method
Gangqinpu applies automated transcription, score alignment, genre labeling, expressive-performance filtering, and joint score grouping to construct a multi-genre aligned piano dataset.
Source: §1 · pp. 1–2Results
Takeaways & Limitations
Gangqinpu provides complementary training data for beat tracking, performance MIDI-to-score transcription, and expressive performance rendering, with cross-dataset gains reported for beat tracking.
Source: §1 · p. 2Beat confidence scores are heuristic annotation-reliability indicators, not calibrated probabilities.
Source: §3.3 · p. 3
Abstract
from arXiv · showhide
Aligned audio, performance MIDI, and musical scores provide valuable supervision for music analysis and performance modeling, yet datasets that jointly align all three modalities remain scarce and limited in scale and genre diversity. We introduce Gangqinpu, a multi-genre piano dataset containing 18,786 score-performance pairs and 1,007.1 hours of source audio, with automatic note-level alignments and derived beat/downbeat annotations. It spans diverse genres, including pop, classical, jazz, folk/traditional, rock, and blues. Our automated pipeline transcribes piano audio, aligns the resulting MIDI with musical scores, assigns genre labels, and filters out non-expressive performances with a score-informed binary classifier. We further construct joint splits with ASAP using conservative score-similarity grouping to reduce data leakage. We evaluate Gangqinpu on performance MIDI-to-score transcription, expressive performance rendering, and beat tracking. Downstream experiments demonstrate the value of Gangqinpu as complementary training data, with cross-dataset gains in beat tracking and primarily in-domain benefits for transcription and expressive performance rendering. The code and dataset are publicly available at https://github.com/wei-zeng98/gangqinpu.
1. INTRODUCTION
Piano datasets rarely align audio, performance MIDI, and scores across diverse genres, limiting the resources available for music analysis and performance modeling. Gangqinpu addresses this gap with a large multi-genre aligned dataset and evaluates its value as complementary training data across three tasks.
1. INTRODUCTION
- Existing piano datasets commonly cover only a subset of audio, performance MIDI, and scores; aligned three-modality datasets remain scarce and often classical-focused.ASAP includes audio for 519 of 1,067 MIDI performances, has some incompletely aligned scores, and is restricted to Western classical piano.Source: §1 · p. 1
- Pop and jazz may use lead sheets, informal arrangements, or improvisation, while expressive-rendering research has focused predominantly on classical piano.The paper identifies transcription from recordings and contemporary production as motivating applications for broader-genre aligned data.Source: §1 · p. 1
- Gangqinpu contains 18,786 score–performance pairs, 1,007.1 hours of audio, and automatic note-level plus beat/downbeat annotations across multiple genres.Named genres include pop, classical, jazz, folk/traditional, rock, and blues.Source: §1 · p. 2
- Its automated pipeline transcribes audio, aligns MIDI to scores, assigns genres, filters non-expressive performances, and jointly groups scores with ASAP to reduce leakage.The dataset is presented as complementary training data for beat tracking, performance MIDI-to-score transcription, and expressive rendering.Source: §1 · pp. 1–2
- Beat tracking improves on both Gangqinpu and ASAP, while transcription and rendering gains are mainly in-domain.The reported gains are for performance MIDI-to-score transcription and expressive performance rendering.Source: §1 · p. 2
2. RELATED WORK
Prior piano datasets vary in capture method, modality coverage, and score alignment, with score-aligned resources predominantly centered on Western classical repertoire. The paper also reviews transcription and alignment systems that inform its pipeline choices.
2.1 Piano Datasets
- MAESTRO offers synchronized captured audio and MIDI, while ATEPP and Aria-MIDI scale transcription from existing recordings but remain subject to transcription errors.Other resources provide varying score and annotation coverage, including note-aligned MIDI collections and beat-level alignments.Source: §2.1 · p. 2
- Gangqinpu addresses the gap between multi-genre piano recordings and independently notated scores by connecting them through transcribed MIDI and automatic note-level alignments.Existing score-aligned resources predominantly cover Western classical repertoire; POP909 has popular-music MIDI annotations but lacks paired piano recordings and independently notated scores.Source: §2.1 · p. 2
2.2 Audio-to-MIDI Piano Transcription and Score–Performance Alignment
- The authors select Transkun v2 for structured note and pedal interval prediction that preserves note offsets without pedal-based extension.This allows note offsets and sustain-pedal activity to be represented separately.Source: §2.2 · p. 2
- DualDTWNoteMatcher separates pitch-sequence matching from pitch-wise onset alignment and reports 99.0% combined note-matching F1 versus 98.5% for the Nakamura baseline.Its reported results and integration with symbolic processing and refinement motivate its use in the pipeline.Source: §2.2 · p. 2
3. DATASET CONSTRUCTION
Gangqinpu is built from online piano audio–score pairs through transcription, symbolic alignment, beat annotation, expressive-performance filtering, genre labeling, and leakage-aware splitting. The construction combines automated procedures with explicit thresholds and heuristic confidence annotations.
3.1 Overall Pipeline
- The pipeline transcribes audio, aligns performance MIDI with score MIDI, refines note correspondences, derives beat/downbeat annotations, filters samples, assigns genres, and groups related scores for joint splits.Duplicate-audio checks and cross-dataset grouping with ASAP are used to reduce leakage.Source: §3.1 · p. 2
3.2 Data Acquisition and Preprocessing
- Collection retains recordings of at least 60 seconds and piano-only scores with at least 32 measures, then converts and cleans score and performance MIDI for alignment.Score preprocessing unfolds repeats and alternative endings; performance tracks are merged, duplicate or implausibly short notes removed, and overlapping same-pitch notes truncated.Source: §3.2 · pp. 2–3
3.3 Alignment-Derived Beat Annotation and Confidence
- Pairs are excluded when raw-alignment insertion or deletion rates exceed 10%; score beats are mapped to performance time using nearby note onsets, with unsupported beats interpolated or extrapolated.Beat confidence is heuristic, not calibrated probability, and is intended to support masking uncertain regions downstream.Source: §3.3 · p. 3
3.4 Expressive Performance Filtering
- A score-informed binary classifier distinguishes expressive from non-expressive performances using eight handcrafted timing and dynamic features derived from observed note correspondences.The labeled subset uses two experts' listening assessments; training also adds ASAP performances as positive examples and synthetic score renderings as negatives, retaining samples with predicted expressive probability of at least 0.5.Source: §3.4 · pp. 3–4
3.5 Genre Annotation and Leakage-Aware Splitting
- Genre labels come first from textual metadata and then, when needed, audio–text comparison; uncertain predictions remain unknown.Score-similarity groups and broader split groups link related scores, arrangements, excerpts, and cross-dataset matches; each group stays within one partition.Source: §3.5 · p. 4
- Gangqinpu and ASAP are jointly grouped for partitioning, with designated ASAP test pieces assigned to test and downstream evaluation reported separately by dataset.The resulting splits contain 17,773/499/514 Gangqinpu samples and 955/50/61 (n)ASAP samples for train/validation/test, respectively.Source: §3.5 · p. 4
4. DATASET STATISTICS
Gangqinpu contains 18,786 paired performances and 1,007.1 hours of source audio, with score-similarity groups used as an operational measure of score diversity. Its genre, duration, and note-count distributions are varied and long-tailed.
4. DATASET STATISTICS
- 18,786 paired performances span 1,007.1 hours of source audio and 15,645 automatically identified score groups.The 2,445 groups with multiple performances contain alternative realizations of similar musical material; score groups are not a verified count of distinct compositions.Source: §4 · p. 4, Table 3 · p. 5
- Pop and classical dominate assigned genre labels, while audio durations and refined-MIDI note counts have long-tailed distributions.Other represented labels include folk/traditional, jazz, rock, blues, other, and unknown; note counts include interpolated notes.Source: §4 · pp. 4–5
5. EXPERIMENTS
Gangqinpu is evaluated as a filtering resource and as complementary training data across beat tracking, performance MIDI-to-score transcription, and expressive rendering. Adding it improves beat tracking on both test datasets, while transcription and rendering gains are mainly within Gangqinpu.
5.1 Expressive Performance Classification
- The eight-feature MLP reaches 98.61% accuracy and 98.41% macro-F1 on 288 expert-labeled held-out samples, and leads pooled accuracy and macro-F1 among the compared classifiers on real data.Its detection rate falls as synthetic onset noise increases; NOMML predicts every sample expressive, yielding 82.58% pooled accuracy and 45.23% macro-F1.Source: §5.1 · p. 5, Table 4 · p. 5
5.2 Downstream Task Evaluation
- Across tasks, ASAP-only and ASAP-plus-Gangqinpu training are evaluated separately on both test datasets using fixed joint splits.The ASAP test sets contain 26 Beat This! samples, 59 PM2S hand-assignment samples, and 61 samples for other PM2S tasks and S2A; Gangqinpu test sets contain 514 samples.Source: §5.2 · p. 5, §5.2.1 · p. 5, §5.2.2 · p. 6
5.2.1 Beat Tracking
- On ASAP, beat and downbeat F1 rise 4.30 and 4.31 percentage points; on Gangqinpu, beat F1 rises 60.18 to 78.16 and downbeat F1 39.60 to 63.66.Adding Gangqinpu improves all reported beat-tracking metrics on both datasets.Source: §5.2.1 · p. 6, Table 5 · p. 6
5.2.2 Performance MIDI-to-Score Transcription
- On Gangqinpu, beat and downbeat F1 rise 14.02 and 20.95 percentage points, and onset and note-value weighted F1 rise 18.87 and 19.32 points.Time-signature macro-F1 declines; transfer to ASAP is mixed, with some attributes improving and others declining.Source: §5.2.2 · p. 6
5.2.3 Expressive Performance Rendering
- On Gangqinpu, velocity MAE falls from 17.50 to 13.05, IOI MAE from 78.51 to 54.66 ms, and duration MAE from 309.47 to 269.04 ms.All three errors increase modestly on ASAP, so the gains do not transfer uniformly across domains.Source: §5.2.3 · p. 6
6. CONCLUSION
Gangqinpu combines 18,786 paired performances and 1,007.1 hours of audio with automatic note-level alignments. Its evaluations show cross-dataset beat-tracking gains, but transcription and rendering benefits are primarily in-domain.
6. CONCLUSION
- Gangqinpu comprises 18,786 paired performances and 1,007.1 hours of source audio, with performance MIDI and scores connected by automatic note-level alignments.Its pipeline combines expressive-performance filtering, confidence-aware beat annotation, genre labeling, and leakage-aware splitting.Source: §6 · p. 6
- Experiments improve beat tracking on Gangqinpu and ASAP, while most transcription targets and all evaluated rendering metrics improve on Gangqinpu; transfer for transcription and rendering remains mixed.The conclusion identifies manual auditing of automatic annotations as future work.Source: §6 · p. 6, Table 6 · p. 7, Table 7 · p. 7
- The conclusion identifies manual auditing of automatic annotations as a focus for future work.Source: §6 · p. 6