Research project

PawCT: End-to-End Choral Transcription into Separate SATB Note Tracks

Hanyu Meng¹, Zhanhong He², Zixun Guo³, Yaolong Ju⁴

¹UNSW · ²UWA · ³Queen Mary University of London · ⁴Great Bay University

Links: [ Paper | Code ]

Most choral transcription systems return one merged stream of notes. PawCT identifies which sections are singing and writes a separate MIDI track for soprano, alto, tenor, and bass.

From one choir recording to four usable parts

A mixed choir is transcribed either into one merged note track or into separate soprano, alto, tenor, and bass tracks. — Conventional choral AMT merges every voice. PawCT preserves the SATB structure needed for rehearsal, education, and score reconstruction.

Performance at a glance

0.382 single-track note F1 +61.18% over the previous state of the art

0.225 average SATB note F1 +36.36% over the adapted choral baseline

0.313 zero-shot note F1 on Cantoría 2.72× the two-stage assignment baseline

Single-track transcriptionNote F1 · YouChorale

Previous SOTA0.237

PagCT0.382

Separate SATB transcriptionAverage note F1 · YouChorale

Adapted choral model0.165

Post-hoc assignment0.175

PawCT-OC0.225

Bold: best. Underlined: second best.

Full paper results

Table 1 · Part-agnostic transcription on YouChorale

Model	#Params	Frame			Onset (50 ms)			Onset (100 ms)
Model	#Params	P	R	F1	P	R	F1	P	R	F1
Onsets and Frames	6.79M	0.806	0.326	0.428	0.450	0.178	0.242	0.688	0.248	0.344
MT3	44.7M	0.590	0.243	0.344	0.117	0.148	0.127	0.200	0.255	0.217
YourMT3+	45.8M	0.529	0.423	0.457	0.145	0.119	0.127	0.252	0.207	0.221
MuScriptor-medium	307M	0.720	0.745	0.730	0.086	0.083	0.084	0.236	0.228	0.231
Yu et al.	4.23M	0.678	0.619	0.647	0.251	0.232	0.237	0.366	0.341	0.347
PagCT w/o Aug.	14.84M	0.690	0.833	0.751	0.346	0.329	0.333	0.539	0.506	0.516
PagCT	14.84M	0.626	0.909	0.738	0.398	0.378	0.382	0.597	0.569	0.573

Table 2 · Part-aware SATB transcription on YouChorale · Note F1 @ 50 ms

Model	Soprano		Alto		Tenor		Bass		Average		VA Rate (%)
Model	Frame	Note	Frame	Note	Frame	Note	Frame	Note	Frame	Note	Frame	Note
YourMT3+^★	0.371	0.083	0.297	0.068	0.294	0.073	0.371	0.093	0.333	0.079	72.87	62.20
Yu et al.^★	0.528	0.183	0.451	0.151	0.436	0.138	0.560	0.186	0.494	0.165	76.35	69.62
PagCT + Post-VA	0.501	0.180	0.439	0.145	0.447	0.156	0.568	0.219	0.489	0.175	66.26	45.81
PawCT w/o union-loss	0.417	0.192	0.338	0.169	0.343	0.178	0.419	0.222	0.379	0.190	51.36	49.74
PawCT	0.530	0.228	0.430	0.191	0.429	0.197	0.552	0.252	0.458	0.217	62.06	56.81
PawCT-RP	0.542	0.210	0.453	0.185	0.458	0.194	0.585	0.247	0.510	0.209	69.11	54.71
PawCT-RP-OC	0.535	0.231	0.447	0.197	0.451	0.208	0.578	0.264	0.503	0.225	68.16	58.90

Table 3 · Cross-dataset part-aware note F1 @ 50 ms

Dataset	Model	Soprano	Alto	Tenor	Bass	Average
CSD n = 3	PagCT + Post-VA	0.073	0.122	0.131	0.143	0.118
	PawCT	0.140	0.175	0.183	0.145	0.172
	PawCT-RP-OC	0.145	0.223	0.199	0.130	0.174
Cantoría n = 14	PagCT + Post-VA	0.179	0.115	0.085	0.084	0.115
	PawCT	0.240	0.239	0.275	0.328	0.271
	PawCT-RP-OC	0.329	0.273	0.300	0.348	0.313

Model Overview

PawCT detects notes and assigns them to vocal parts in one model. A global training signal keeps the complete pitch content intact, while pitch range and melodic continuity help maintain consistent SATB lines.

PawCT predicts separate SATB note events directly, while the comparison systems produce one merged track or assign parts after transcription. — PawCT predicts SATB notes directly from the mixed recording. The two-stage pipeline is shown only for comparison: it first transcribes a merged track and assigns parts afterwards, and is not part of PawCT.

Listen to the transcription

This 35-second excerpt of Exsultate Deo lets you compare the part-agnostic ground truth with merged predictions from our PagCT model, MuScriptor-medium, and Yu & Duan (2024). PawCT additionally returns four independently playable vocal parts.

Original recording Palestrina · Exsultate Deo · 00:10–00:45

35-second on-page excerpt

Preparing waveform…

0:00 / 0:35

The audio stays on this page—click the waveform to seek. Recording source and attribution

Loading the reference and model outputs…

Merged transcription

All annotated or detected notes share one track.

Ground truth Reference · merged note track

F1@50 ms: —
F1@100 ms: —

PagCT Ours Our part-agnostic model · single note track

F1@50 ms: 0.433
F1@100 ms: 0.668

MuScriptor-medium [1] Rouard et al., 2026

F1@50 ms: 0.037
F1@100 ms: 0.240

Yu & Duan (2024) [2] ISMIR 2024

F1@50 ms: 0.017
F1@100 ms: 0.030

Why can MuScriptor still look plausible?

A 35-second overview compresses a 50–100 ms onset shift to roughly one or two pixels. Note F1 requires an exact-pitch, one-to-one match whose onset falls inside the stated tolerance, so broadly correct pitch contours can coexist with low event scores. MuScriptor's rise from 0.037 at 50 ms to 0.240 at 100 ms is consistent with many timing near-misses; the remaining gap also permits missed, extra, split, or merged notes.

What does the comparison establish?

PagCT is task-specific, whereas MuScriptor-medium uses the released general-purpose checkpoint. This excerpt therefore illustrates task-aligned event localization versus out-of-domain generalization—not an architecture comparison under matched training. PagCT is substantially better under this event metric, but its 0.433 F1 also makes clear that the transcription is not error-free. The full YouChorale test-set results in the paper remain the primary quantitative evidence.

Separate vocal parts

Play each PawCT output on its own, or compare it with post-hoc assignment.

SopranoS

AltoA

TenorT

BassB

View the full piano-roll comparison

Ground truth, PawCT, PagCT, and PagCT plus Post-VA piano rolls for Exsultate Deo. — Ground truth and model outputs for the 10–100 second passage. Color denotes the assigned SATB part.

References

S. Rouard, M. Krause, A. Roebel, C.-J. Simon-Gabriel, and A. Défossez, “MuScriptor: An Open Model for Multi-Instrument Music Transcription,” arXiv:2607.08168 [cs.SD], 2026. Paper · Code · Model
H. Yu and Z. Duan, “Note-Level Transcription of Choral Music,” in Proceedings of the 25th International Society for Music Information Retrieval Conference (ISMIR), pp. 182–188, 2024. Paper · Code

Error analysis

From pitch activity to note events

MuScriptor-medium and PagCT achieve similar frame F1 (0.730 vs. 0.738), but PagCT obtains substantially higher 50-ms note F1 (0.382 vs. 0.084). This suggests that the main gap lies in localizing note events precisely rather than detecting sustained pitch activity.

Comparison of frame F1 and note F1 for all baselines, highlighting MuScriptor's large frame-to-event drop. — Strong frame-level pitch tracking does not necessarily translate into accurately localized note events.