AI Lofi HipHop - How (Not) To Generate AI Music
summary
A 24/7 live stream plays endlessly generated lofi hip‑hop produced by a custom pipeline that uses Magenta MusicV + Groove A + a transformer multitrack model to make MIDI, Python synths (FluidSynth, a sample‑based flute synth and a hybrid synth), a Pyomix recipe engine, and ffmpeg + Flask to push video to YouTube. Training needs (MusicV used 170,000 MIDIs), scarce lofi MIDI data, a broken tape emulator, and manual mixing hacks mean the system runs continuously but doesn't yet sound convincingly lofi.
AI-written summary · about
chapters (14)
- 0:00Welcome to AI Lofi HipHop
- 6:08Building our AI music generator
- 12:29Challenges in direct audio AI
- 14:01Sheet music with Magenta Music VAE
- 21:03Drum loops with Groove A
- 23:09Melodies with Multitrack Machine
- 25:37Sheet music to audio with synthesizers
- 31:23Mixing and effects
- 34:37Streaming to YouTube
- 38:33Live demo of AI Lofi stream
- 39:58Why AI Lofi sounds off
- 43:55Future improvements
- 48:38Introducing Pyomix V2
- 50:27Q&A
related talks
45mAI Lofi HipHop - How Not To Generate AI-Music
41mMusic and Machine Learning
36mMusic and Art with ML
46mMusic Generation with Magenta: Using Machine Learning in
01h 03mAI Hits Different - A Music Album From the Machine Room
24mCreating an AI Musician with JavaScript
28mMusic and AI in the Browser with TensorFlow js and Magenta js
34mHooray for Side Projects! Smooth Music for Coding, Driven by Machine Learning