PawCT and PagCT
Two PyTorch checkpoints for transcribing mixed choral audio to note-level MIDI. PagCT produces one merged note track. PawCT base produces separately labelled soprano, alto, tenor, and bass tracks using the end-to-end model with union supervision.
| Model | Checkpoint | Inference config | YouChorale note F1 @ 50 ms | Regression thresholds (frame / onset / offset) |
|---|---|---|---|---|
| PagCT | pagct.pth | pagct.json | 0.3790 | 0.07 / 0.03 / 0.003 |
| PawCT base | pawct-base.pth | pawct-base.json | 0.2248 | 0.05 / 0.01 / 0.01 |
Transcribe audio
Use the Paw Choral AMT code at revision b0718f8 and its documented dependencies. Download the checkpoint and matching JSON into the same folder and run the example from that folder. Set checkout to your code checkout and audio_path to a choral audio file.
import json
import sys
from pathlib import Path
import librosa
from omegaconf import OmegaConf
checkout = Path("/path/to/Paw_Choral_AMT")
audio_path = Path("/path/to/choir.wav")
model_id = "pawct-base" # Use "pagct" for a single merged track.
sys.path.insert(0, str(checkout / "src"))
from inference import ChoralAMTTranscriber
from infer_midi_voice_assignment import write_satb_midi
settings = json.loads(Path(f"{model_id}.json").read_text(encoding="utf-8"))
cfg = OmegaConf.merge(
OmegaConf.load(checkout / "src" / "config.yaml"), settings["public_code_overrides"]
)
audio, _ = librosa.load(audio_path, sr=cfg.feature.sample_rate, mono=True)
transcriber = ChoralAMTTranscriber(cfg, settings["checkpoint_file"])
output_midi = Path(f"{model_id}.mid")
if cfg.choral.enable:
probabilities = transcriber.transcribe(audio)["output_dict"]
tracks = {}
for index, voice in enumerate(cfg.choral.voice_names):
rolls = {
f"{key}_output": probabilities[f"voice_{key}_output"][:, index, :].copy()
for key in ("frame", "onset", "offset")
}
tracks[voice], _ = transcriber.post_processor.output_dict_to_midi_events(rolls)
write_satb_midi(tracks, output_midi)
else:
transcriber.transcribe(audio, midi_path=str(output_midi))
PawCT uses the same thresholds for each voice and applies its learned presence gate. The bundled configurations retain the compatibility flags needed by the legacy PyTorch checkpoints.
Evaluation and provenance
The table reports a re-score of archived predictions on the official 30-recording YouChorale test split. Note matching requires the correct pitch and an onset within 50 ms. The metric ignores offsets. Scores average over recordings, and PawCT also averages across the four SATB voices.
The archived probabilities lack a checkpoint hash. The threshold history includes test-set selection, so treat these scores as descriptive results for the published settings. See evaluation details, checkpoint provenance, and file checksums.
Citation
Hanyu Meng, Zhanhong He, Zixun Guo, and Yaolong Ju. Toward Part-Aware Choral Transcription with Singing Voice Assignment. 2026. arXiv:2610.09861