A Composer’s Guide to Creating with Generative Neural Networks
summary
An independent composer trains WaveNet (and later Dance Diffusion) from scratch on 35–50 minutes per performer augmented to ~8–9 hours, runs ~300–400k training steps per model over weeks, spends several hundred dollars per run and builds a sampler that outputs 1–4s timbral snippets which are stitched into a multi-movement piece for accordion and percussion. Challenges include data augmentation artifacts, multimodal instrument training, slow convergence and infrastructure choice; results are musically usable proof-of-concept samples and a working live-electronics system.
AI-written summary · about
chapters (14)
- 0:00Intro
- 4:53What's in this presentation, and what not?
- 5:43Outline
- 6:05Background & goals
- 13:22Data collection
- 18:43Model selection
- 25:05Infrastructure
- 25:47Development
- 29:22Deciding when it's done
- 34:09Composition
- 37:24Future directions
- 40:45Creative exercise
- 41:50Gratitude
- 42:10Outro
related talks
32mGetting started with generative AI with AWS DeepComposer
30mSound Modelling - parametric methods and deep learning...
01h 04mKeynote 3. Interactive generative models for music
46mMusic Generation with Magenta: Using Machine Learning in
21mMusical Deep Neural Networks in the Browser
34mCreativity and AI: Deep Neural Nets "Going Wild"
24mCreating an AI Musician with JavaScript
51mComputational Creativity