VioPTT: Violin Playing Technique-Aware
Transcription from Synthetic Data Augmentation

Abstract

While automatic music transcription is well-established in music information retrieval, most models are limited to transcribing pitch and timing information from audio, and thus omit crucial expressive and instrument-specific nuances. One example is playing technique on the violin, which affords its distinct palette of timbres for maximal emotional impact.

Here, we propose VioPTT (Violin Playing Technique-aware Transcription), a lightweight cascade model that directly transcribes violin playing technique in addition to pitch onset and offset. Furthermore, we release MOSA-VPT, a novel, high-quality synthetic violin playing technique dataset to circumvent the need for manually labeled annotations. Leveraging this dataset, our model demonstrated strong generalization to real-world note-level violin technique recordings in addition to achieving state-of-the-art transcription performance. To our knowledge, VioPTT is the first to jointly combine violin transcription and playing technique prediction within a unified framework.

System Overview

VioPTT comprises two modules: a transcription module that predicts note onset, offset, velocity, and frame activation at the frame level; and an articulation module that fuses per-note acoustic embeddings with transcription features to classify playing technique.

Playing Techniques

MOSA-VPT Dataset

We introduce MOSA-VPT, a large-scale synthetic dataset of 76 hours of audio–MIDI pairs balanced across four techniques: détaché, flageolet, spiccato, and pizzicato. Audio is rendered from MOSA MIDI scores using DAWDreamer with a professional virtual violin instrument (Synchron Solo Violin I), with all room and spatial processing disabled. The dataset is fully annotation-free.

→ Download MOSA-VPT on Zenodo

Audio Examples

MOSA-VPT Synthetic Dataset Samples

Samples from our synthetic MOSA-VPT dataset — violin audio rendered from MOSA MIDI scores using a professional virtual violin instrument (Synchron Solo Violin I) via DAWDreamer. Each track is a 30-second excerpt rendered with a different playing technique. The dataset contains 76 hours of audio across four techniques.

Results

Pitch & Timing Transcription (Table 1)

Transcription performance on URMP and Bach10 violin tracks (Precision / Recall / F1 / onset-only F1). Our model trained from scratch with augmentation matches state-of-the-art MUSC on URMP while outperforming it on Bach10 despite using ~30% less training data.

Playing Technique Classification — Ablation Study (Table 2)

Macro accuracy and per-class accuracy on the RWC dataset (mean ± std over 3-fold cross-validation). The no ablation condition (all four transcription features) achieves the best macro accuracy of 77.22%, outperforming MERTech (53.36%) by a large margin.

Citation

If you find this work useful, please cite:

@article{wang2025vioptt,
  title   = {VioPTT: Violin Playing Technique-Aware Transcription from Synthetic Data Augmentation},
  author  = {Wang, Ting-Kang and Peng, Yueh-Po and Su, Li and Cheung, Vincent K.M.},
  journal = {arXiv preprint arXiv:2509.23759},
  year    = {2025}
}