How to Use Machine Learning and .NET to Write Crushing Metal Riffs
summary
Hands-on experiment trains a sequence model on roughly 200 short riffs from Metallica, Slayer and Anthrax converted to a six-string fret/time token format, using stacked LSTMs with 0.3 dropout for sequence generation, about ~2M parameters trained for ~200 epochs (~1 hour) in Python Keras/TensorFlow, then exported attempts via ONNX/CNTK/ML.NET back into .NET. Output transcribes to Guitar Pro, reproduces power‑chord patterns and bar timing but sounds rough, showing promising but imperfect riff generation and clear routes for improved data, architecture or automated export.
AI-written summary · about
chapters (20)
- 0:00Welcome and introduction
- 1:31What is machine learning
- 3:18Neural networks basics
- 6:26Recurrent neural networks
- 9:19Generating metal music
- 10:27Representing guitar riffs
- 11:32Data preparation
- 13:02ML.NET for .NET developers
- 16:40TensorFlow.NET alternatives
- 18:50Python experiment setup
- 19:12Training the model
- 22:29Generating the riff
- 23:32Live demo of AI riff
- 24:43Improving the results
- 25:05Exporting to .NET
- 26:40ML.NET for simple tasks
- 27:21Advanced .NET ML frameworks
- 28:24Future improvements
- 29:51Generative adversarial networks
- 30:15Conclusion and Q&A
