Phrase-Level Modeling of Expression in Violin Performances

Area of Science:

• Music Performance and Pedagogy
• Computational Musicology
• Machine Learning in Music

Background:

• Musical expression is vital but challenging to teach.
• Computational models offer potential for feedback in music lessons.
• Expert performances can inform models for student guidance.

Purpose of the Study:

• To model expressive variations in dynamics and timing for solo violin.
• To facilitate learning expressive performance in new, unrecorded pieces.
• To analyze phrase-level predictions against expert performances.

Main Methods:

• Phrase-level prediction model based on musical score and idiomatic expressiveness.
• Evaluation using numerical datasets and comparison with note-level models.
• Perceptual evaluation by listeners comparing synthesized and expert performances.

Main Results:

• Listeners did not consistently identify synthesized performances as human-like.
• Exclusion of features like articulation and vibrato may impact perception.
• Some listeners provided positive feedback on modeled renditions.
• Phrase-level modeling showed slight numerical degradation but improved visual interpretability.

Conclusions:

• Current phrase-level modeling may not fully capture perceived musical expression.
• Articulation and vibrato are potentially critical for realistic violin performance.
• The model shows promise for pedagogical applications due to interpretable predictions.
• Further research should incorporate a wider range of expressive features.