Source-linked AI summary
Toward Part-Aware Choral Transcription with singing voice assignment
Hanyu Meng, Zhanhong He, Zixun Guo, Yaolong Ju
TL;DR
Recent note-level choral AMT merges concurrent SATB voices rather than transcribing them as separate parts. PawCT jointly identifies active parts and transcribes separate note-level tracks; on YouChorale, it improves part-aware and part-agnostic note F1 over prior alternatives.
Problem
Recent note-level choral AMT produces a single merged transcription without vocal-part labels, while similar timbres and overlapping pitch ranges make SATB assignment ambiguous.
Source: §1 · p. 1Method
PawCT jointly predicts SATB-specific note events and active-part presence, using union-level supervision plus range-prior and ordered-continuity targets.
Source: §1 · p. 1Results
PawCT-RP-OC achieves macro part-aware note F1 of 0.225 versus 0.165 for the adapted choral model and 0.175 for PagCT+Post-VA; PagCT reaches 0.382 50-ms onset F1 versus 0.237 for the previous choral model.
Source: §4.3 · p. 4, §4.2 · p. 4Takeaways & Limitations
The results support jointly modeling note transcription and vocal-part assignment, with CSD and Cantoría evaluations further assessing performance under dataset shift.
Source: §5 · p. 4, §4.4 · p. 4
Abstract
from arXiv · showhide
Single-instrument automatic music transcription (AMT) has advanced substantially, yet choral applications require soprano, alto, tenor, and bass (SATB) to be transcribed as separate parts. Recent note-level choral AMT instead produces a single merged note track, limiting rehearsal, education, and score reconstruction. To address this limitation, we introduce Part-aware Choral Transcription (PawCT), to our knowledge the first end-to-end neural framework that identifies active SATB parts from choral audio and transcribes each into a separate note-level track. PawCT combines part-specific onset, offset, and frame prediction with part-presence estimation, union-level supervision, and structured training targets using a range prior (RP) based on SATB pitch ranges and its ordered-continuity (OC) extension, which adds within-part melodic continuity and cross-part pitch ordering. On YouChorale, PawCT-RP-OC achieves a macro part-aware note F1 of 0.225 at a 50-ms onset tolerance, outperforming an adapted choral baseline (0.165) by 36.4% relative and a two-stage post-hoc assignment pipeline (0.175). Its part-agnostic variant, PagCT, achieves a 50-ms onset F1 of 0.382, compared with 0.237 for the previous state-of-the-art choral AMT model. Cross-dataset evaluations on CSD and Cantoria further assess performance under dataset shift. These results demonstrate the benefit of jointly modeling note transcription and vocal-part assignment. Code and demos are available at https://hanyu-meng.github.io/Paw_Choral_AMT_Demo/.
1. INTRODUCTION
Recent note-level choral AMT produces a merged transcription without part labels, despite the need to distinguish concurrent SATB parts. PawCT jointly assigns active parts and transcribes each into a separate note-level track.
1. INTRODUCTION
- AMT has advanced across single-instrument, ensemble, and singing-voice tasks, but choral singing remains underexplored.The introduction notes rapid progress in piano transcription through onset-conditioned and high-resolution note models and paired audio–MIDI datasets.Source: §1 · p. 1
- Recent note-level choral AMT produces a single merged transcription without vocal-part labels, while similar timbres and overlapping pitch ranges make assignment ambiguous.Earlier approaches focused on pitch-level trajectories or streams rather than explicit note-event boundaries.Source: §1 · p. 1
- On YouChorale, PawCT-RP-OC outperforms adapted and two-stage baselines in part-aware note F1, while PagCT improves part-agnostic note F1.The framework also reports gains on CSD and Cantoría over the two-stage baseline.Source: §1 · p. 1
- PawCT combines SATB-specific onset, offset, frame, and presence prediction with union-level supervision and range-prior and ordered-continuity training targets.These contributions support joint note transcription and active-part assignment from mixed choral audio.Source: §1 · p. 1
2. METHODS
PawCT jointly predicts SATB-specific note activity and active parts, while a two-stage alternative decodes a merged PagCT track before symbolic Post-VA assignment.
2. METHODS
- PawCT jointly predicts SATB-specific note activity from audio, while PagCT plus Post-VA first decodes a merged note track and then assigns its notes to vocal parts.The paper compares end-to-end joint transcription with post-hoc symbolic voice assignment.Source: §2 · pp. 1–2, Figure 2 · p. 2
2.1. Part-aware Choral Transcription Architecture
- A shared CRNN encodes the audio; part-specific heads estimate onset, offset, and frame activity, while a presence head gates each part's predicted rolls.The encoder uses convolutional blocks and a bidirectional GRU; the presence head estimates active parts.Source: §2 · p. 1, §2.1 · p. 2
2.2. Voice-Assignment Priors
- The range prior assigns notes using SATB register, and ordered continuity adds melodic-continuity and pitch-order constraints to assignments.For onset groups of up to four notes, OC selects one-to-one assignments preserving pitch order; larger groups use independent minimum-cost assignment.Source: §2.2 · p. 2
2.3. PawCT Training Objective
- PawCT training combines masked part-wise transcription loss, part-presence BCE, and union-level supervision against targets built from merged ground-truth notes.The union loss preserves global pitch content by comparing pooled part predictions with merged-note targets.Source: §2.3 · pp. 2–3
2.4. Part-agnostic Choral Transcription and Post-VA
- PagCT predicts a single global note track without SATB-specific or presence heads; Post-VA classifies decoded notes into SATB parts with an auxiliary active-part head.Post-VA also uses pitch-range and within-onset crossing regularizers.Source: §2.4 · p. 3
2.5. Note Event Decoding
- Notes are decoded from onset, offset, and frame predictions; thresholds are selected on YouChorale validation data and fixed for testing across datasets.The decoder starts notes at onset peaks and ends them when offset activity exceeds threshold, frame activity falls below threshold, or another onset occurs.Source: §2.5 · p. 3
3. EXPERIMENTAL SETUPS
The experiments use YouChorale and pitch-transposition augmentation, compare part-agnostic and part-aware systems, and train PawCT variants with fixed optimization settings.
3.1. Data
- YouChorale provides 22 hours and 25 minutes of a cappella recordings across 452 recordings from 261 songs, with manually aligned MIDI annotations.Most recordings contain all four SATB parts (418; 92.48%); the remaining 34 contain subsets.Source: §3.1 · p. 3
- Pitch-transposition augmentation expands the YouChorale training set sevenfold to approximately 129 hours and 38 minutes of audio.Mixtures and corresponding MIDI and note annotations are jointly transposed by one to three semitones in either direction.Source: §3.1 · p. 3
3.2. Model Configurations
- Part-agnostic comparisons include PagCT and established AMT baselines, while part-aware experiments compare PawCT variants, adapted models, and PagCT + Post-VA.The part-aware ablations test union loss, RP, and OC; the cross-dataset table evaluates models trained only on YouChorale.Source: §3.2 · pp. 3–4, Table 3 · p. 4
3.3. Training Settings
- PawCT and PagCT use AdamW with batch size 8, initial learning rate 10^-4, and early stopping; Post-VA trains for up to 80 epochs.PawCT and PagCT train for up to 200k iterations, with validation-based checkpoint selection.Source: §3.3 · p. 4
4. RESULTS AND DISCUSSION
PawCT-RP-OC leads the part-aware comparisons on YouChorale, while PagCT improves part-agnostic note transcription. Cross-dataset tests show differing performance across CSD and Cantoría.
4.1. Evaluation Metrics
- Evaluation reports part-agnostic note F1 at 50- and 100-ms onset tolerances and part-aware macro note F1 at 50 ms.VA Rate is the percentage of part-agnostic F1 retained by macro-average SATB F1.Source: §4.1 · p. 4
4.2. Part-agnostic Transcription
- 0.382 50-ms onset note F1 for augmented PagCT is 61.2% above the previous choral model's 0.237.Onsets and Frames has high precision but low recall, while MuScriptor's frame performance does not carry over to note F1.Source: §4.2 · p. 4
4.3. Part-aware Transcription
- 0.225 macro part-aware note F1 for PawCT-RP-OC exceeds 0.165 for the adapted choral model by 36.4% and 0.175 for PagCT + Post-VA.PawCT-RP-OC also achieves a note-level VA Rate of 58.90%, compared with 45.81% for PagCT + Post-VA.Source: §4.3 · p. 4
- Union supervision raises macro note F1 from 0.190 to 0.217; RP reaches 0.510 frame F1 but lowers note F1 to 0.209, and OC raises note F1 to 0.225.The results indicate that range guidance alone does not resolve note assignments in overlapping vocal ranges, while OC adds melodic continuity and pitch ordering.Source: §4.3 · p. 4
4.4. Cross-Dataset Evaluation
- On CSD, PawCT and PawCT-RP-OC score 0.161 and 0.174 average note F1 versus 0.118 for the two-stage baseline; on Cantoría, they score 0.271 and 0.313 versus 0.115.Compared with YouChorale, end-to-end models perform worse on CSD but better on Cantoría; the authors suggest dataset structure, recording conditions, and annotation protocols may contribute.Source: §4.4 · p. 4
5. CONCLUSION
The conclusion presents PawCT as an end-to-end approach to separate SATB transcription and reports gains in both part-agnostic and part-aware YouChorale results.
5. CONCLUSION
- PawCT identifies active SATB parts and transcribes each into a separate note-level track.Its framework combines part-specific prediction, union-level supervision, and range-prior and ordered-continuity targets.Source: §5 · p. 4
- PagCT improves part-agnostic note F1 from 0.237 to 0.382 on YouChorale.PawCT-RP-OC achieves the highest macro part-aware note F1 among the evaluated systems and outperforms post-hoc voice assignment.Source: §5 · p. 4
- The reported results support jointly modeling note transcription and vocal-part assignment.Experiments on CSD and Cantoría additionally assess generalization under dataset shift.Source: §5 · p. 4