Music Generation with Magenta: Using Machine Learning in
summary
Magenta (TensorFlow) demos include encoding two WAVs into a WaveNet-autoencoder latent space, summing the vectors and decoding to 1s audio (16 kHz) samples, and sequencing those samples via MIDI in Ableton Live. Prebuilt models used are drums_rnn and melody_rnn with adjustable temperature (e.g. 1.2) to control randomness, plus VAE and LSTM explanations for interpolation and memory. A live app records MIDI clips, writes MIDI files, and uses magenta.js / Magenta Studio; WaveNet sample-by-sample synthesis is slow but NSynth/consent ports make it practical.
AI-written summary · about
chapters (27)
- 0:00Introduction
- 0:27Agenda
- 0:39Definition
- 0:55Making music without being a musician
- 1:23Monica Dean Coolidge
- 2:37Brian Eno
- 3:23Machine Learning in Music Generation
- 3:59MIDI
- 4:49Audio
- 5:23Network Architecture
- 6:09LongTerm Memory
- 7:05Variational Auto Encoder
- 8:35Wavenet
- 9:37Network List
- 10:13Live Code
- 10:42Sound Selection
- 11:44Coding
- 15:45Consent
- 16:28Results
- 17:58Sequence
- 24:42Ableton Sound Synthesis
- 33:07Recap
- 40:43Magenta Jazz
- 42:09Magenta Gain Harp
- 42:48Using Magenta JS
- 43:50Using Magenta Studio
- 44:35Projects
related talks
41mMusic and Machine Learning
47mUnleashing your musical creativity with Magenta.js
36mMusic and Art with ML
28mMusic and AI in the Browser with TensorFlow js and Magenta js
21mMusical Deep Neural Networks in the Browser
01h 04mKeynote 3. Interactive generative models for music
24mCreating an AI Musician with JavaScript
47mExpressive AI-Generated Music With Google's Performance RNN