STRAdi

KO | EN

gitlite — search

#claude #api #docker #css #php #anthropic #large-language-models #deep-learning #animation #devops

(★ 9)

Official code and pretrained models for Real-time Violin Transcription for Score Following (ISMIR 2026)

File Explorer

ZIP Download

- README.md
- example.csv
- example.wav
- manifest.csv
- README.md
- - __init__.py
  - offline.py
  - online.py
- __init__.py
- checkpoint.py
- constants.py
- decoding.py
- inference.py
.gitignore
CITATION.cff
demo.ipynb
evaluate.py
LICENSE
README.md
requirements.txt
transcribe.py

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Official transcription code and pretrained models for our ISMIR 2026 paper,
**STRAdi: Real-time Violin Transcription for Score Following**.

STRAdi (Streaming Transcription for Real-time Alignment) converts solo violin audio into note events and MIDI.
The main model is fully causal and uses no future audio context.
We also release the matched non-causal model used as the offline reference in the paper.

## Models

Both released models are trained on the accessible solo-violin audio released
with MUSC and the violin portion of MOSA.

| Checkpoint | Mode | Training data |
| --- | --- | --- |
| [`violin_transcription_online.pth`](https://github.com/seingreen/real-time-violin-transcription/releases/download/v1.0.0/violin_transcription_online.pth) | Fully causal | MUSC + MOSA |
| [`violin_transcription_offline.pth`](https://github.com/seingreen/real-time-violin-transcription/releases/download/v1.0.0/violin_transcription_offline.pth) | Non-causal | MUSC + MOSA |

Download the checkpoints from
[Release v1.0.0](https://github.com/seingreen/real-time-violin-transcription/releases/tag/v1.0.0)
and place them in `checkpoints/`. The weights are release assets and are not
stored in Git.

## Installation

The paper experiments used Python 3.7.12 and PyTorch 1.13.1.

```bash
conda create -n violin-transcription python=3.7.12 -y
conda activate violin-transcription
pip install -r requirements.txt
```

For a CUDA installation, install the matching PyTorch build for your system
before installing the remaining requirements.

## Transcription

Run the main online model on a solo violin recording.

```bash
python transcribe.py path/to/violin.wav \
  --checkpoint checkpoints/violin_transcription_online.pth \
  --model online \
  --output-dir outputs
```

The command writes a MIDI file and a CSV file with `onset`, `offset`, and
`midi_number` columns. Input audio is converted to mono and resampled to
16 kHz. The onset, frame, and offset thresholds are 0.5 by default.

Use the offline reference by changing the checkpoint and model mode.

```bash
python transcribe.py path/to/violin.wav \
  --checkpoint checkpoints/violin_transcription_offline.pth \
  --model offline \
  --output-dir outputs
```

`demo.ipynb` provides the same online transcription workflow in a notebook.
The repository includes a 60-second KRAISLER excerpt under `examples/`.

## Evaluation

Evaluation data are not included. Prepare a CSV manifest with one row per
audio file.

```csv
id,dataset,audio,annotation
piece01,my_dataset,audio/piece01.wav,annotations/piece01.csv
piece02,my_dataset,audio/piece02.wav,annotations/piece02.csv
```

Paths may be absolute or relative to the manifest. Each annotation CSV must
contain onset and offset times in seconds and a MIDI pitch.

```csv
onset,offset,midi_number
0.520,1.130,69
1.240,1.810,71
```

Run note-level evaluation with the paper settings.

```bash
python evaluate.py \
  --checkpoint checkpoints/violin_transcription_online.pth \
  --model online \
  --manifest path/to/manifest.csv
```

The evaluator reports macro-averaged precision, recall, F1 with onset and
offset, and onset-only F1. It uses a 50 ms onset tolerance, a 20% offset
ratio, and 50-cent pitch tolerance.

The included excerpt can be used for a quick evaluation check.

```bash
python evaluate.py \
  --checkpoint checkpoints/violin_transcription_online.pth \
  --model online \
  --manifest examples/manifest.csv
```

## Repository Scope

This repository contains the transcription models, inference code,
transcription evaluation code, and one short example. It does not distribute
the full training or evaluation datasets or the score-following
implementation.

## Acknowledgments

The architecture adapts the CRNN transcription model used by
[VioPTT](https://github.com/y10ab1/VioPTT), which builds on the
[ByteDance piano transcription code](https://github.com/bytedance/piano_transcription).

## Citation

```bibtex
@inproceedings{lee2026realtime,
  author    = {Sein Lee and Jiyun Park and Juhan Nam},
  title     = {Real-time Violin Transcription for Score Following},
  booktitle = {Proceedings of the International Society for Music Information
               Retrieval Conference (ISMIR)},
  year      = {2026}
}
```

## License

The source code is released under the Apache License 2.0. The example audio
and annotations are subject to the KRAISLER dataset terms.

STRAdi

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README