bowdet 0.2.1

Audio-based bow-change and note-boundary detection for string instruments

bowdet

Audio-based bow-change and note-boundary detection for string instruments.

bowdet detects bow changes (down-bow ↔ up-bow transitions) in bowed string instrument recordings using a two-stage mel-spectrogram boundary approach.

Version 0.2.1 also includes BowDet-NB, an experimental unsupervised note-boundary detector for bowed string recordings.

Install

pip install bowdet

Bow-change detection

from bowdet import detect

times = detect("recording.wav")
print(times)  # array of bow-change times in seconds

Options

times = detect(
    "recording.wav",
    threshold=0.40,
    min_dist_sec=0.12,
)

BowDet-NB: note-boundary detection

BowDet-NB detects note boundaries using an unsupervised V3 multi-scale mel boundary-strength method. It uses only the boundary proposal stage, without the supervised Stage-2 classifier.

from bowdet import detect_nb

times = detect_nb("recording.wav")
print(times)  # array of note-boundary times in seconds

The full descriptive API is also available:

from bowdet import detect_note_boundaries

times = detect_note_boundaries("recording.wav")

Method

The bow-change detector uses a two-stage pipeline:

1. Stage 1 — Boundary Proposal: Computes log-mel spectrogram boundary strength and proposes candidate positions at spectral peaks.
2. Stage 2 — Candidate Classification: A compact CNN fused with 15 acoustic boundary features classifies each candidate as a bow change or not.

BowDet-NB uses a standalone unsupervised version of Stage 1 for note-boundary detection, based on multi-scale left-right mel-spectrogram differences.

Key insight: Bow changes and bowed-string note boundaries often produce broadband spectral transitions visible in the log-mel spectrogram, spanning approximately 100–300 ms across frequency bands.

Performance

Bow-change detection

Evaluated under leave-one-performer-out (LOPO) on 6 viola performers (9 recordings, ~17.7 min, 1,020 annotated bow changes):

Metric	Score
IoU@0.1 F1	0.645 ± 0.012
Point F1 @100ms	0.635

BowDet-NB note-boundary detection

Evaluated on 9 bowed-string recordings spanning viola, violin, and cello:

Metric	Tonal mean F1
Point F1 @50ms	0.701
Point F1 @100ms	0.774
Point F1 @150ms	0.793
Pseudo-region IoU@0.1 F1	0.812
Pseudo-region IoU@0.3 F1	0.803

Citation

If you use bowdet in your research, please cite this repository:

@software{yuan2026bowdet,
  author = {Haotian Yuan},
  title  = {bowdet: Audio-based bow-change and note-boundary detection for string instruments},
  year   = {2026},
  url    = {https://github.com/Yuan-618/bowdet}
}

License

MIT