Modelling emotional nuance of musical performance
summary
Claims a pipeline that models piano nuance from MIDI, using Maestro dataset plus private practice recordings to reach ~200 million tokens and training GPT-2 on 200M tokens. Shows tokenization/quantization choices (event-based tokens, exponential time bins, velocity buckets), removes pedals for initial experiments, and trains ~20 models at ~12 GPU-hours each. Presents generation that mostly preserves key and tempo but is “off yet promising,” shares code on GitHub and data on Hugging Face, and reports concrete augmentation and evaluation ideas.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 3:02Math and music history
- 8:40Piano performance as data
- 14:01Emotional nuances in interpretation
- 20:47Data augmentation strategies
- 25:53Training with MIDI
- 29:01Future plans
- 32:11Conclusions
related talks
33mContext-Aware Masking for Controlled Music Generation in MIDI
47mExpressive AI-Generated Music With Google's Performance RNN
41mMusic and Machine Learning
17mPiano Genie
01h 04mKeynote 3. Interactive generative models for music
30mSound Modelling - parametric methods and deep learning...
58mMusic Generation via Masked Acoustic Token Modeling
30mGenerating Music From Emotion (and other experiments)