Multimodal Violin Technique Recognition with Audio and Skeleton Features

Automatic analysis of violin performance can support music education by helping students and teachers identify playing techniques from ordinary practice videos. This paper focuses on violin technique recognition rather than full automated grading. A simple multimodal framework is used to extract audio descriptors from the soundtrack and skeleton-motion descriptors from video frames, then combine the two modalities for single-label technique classification. The current study considers six dominant technique categories: pizzicato, stopped bowing, plain playing, vibrato, double stops, and spiccato. Experiments on 147 labeled clips show that multimodal fusion achieves 96.60% accuracy and 96.43% macro-F1 under stratified clip-level three-fold cross-validation. A source-video-grouped audit gives 95.92% accuracy and 96.08% macro-F1. The results provide preliminary evidence that audio and skeleton cues are complementary for violin technique recognition.