MuQ Piano Performance Evaluation Model
Audio-based piano performance evaluation using MuQ layers 9-12 with Pianoteq soundfont ensemble training.
Model Details
- Architecture: MuQ L9-12 with mean+std pooling -> 512 hidden -> 19 dimensions
- Training: 3-fold cross-validation with 6 Pianoteq soundfonts for augmentation
- Performance: R² = 0.537 (55% improvement over symbolic baseline)
- Paper: Audio Foundation Models Outperform Symbolic Representations for Piano Performance Evaluation
Soundfonts Used for Training
- HB Steinway Model D (bright concert grand)
- YC5 Vintage (balanced)
- K2 Basic (warm)
- NY Steinway D Honky Tonk
- NY Steinway D Worn Out
- U4 Small (intimate upright)
19 Perceptual Dimensions
Timing, Articulation (length, touch), Pedal (amount, clarity), Timbre (variety, depth, brightness, loudness), Dynamics, Tempo, Space, Balance, Drama, Mood (valence, energy, imagination), Interpretation (sophistication, overall)
Usage
This model is deployed as a HuggingFace Inference Endpoint. See the handler.py for the inference API.
License
Apache 2.0
