notlame-train

Train a neural network to replace LAME's psychoacoustic model for MP3 encoding.

Quick Start

make setup              # Install dependencies
make download-gtzan     # Download music (1.2GB)
make prepare            # Convert to MDCT
make train              # Train model
make test               # Run 141 tests (verify code)
make evaluate           # Compare vs LAME
make export             # Export to ONNX

What It Does

Traditional MP3:  Audio → LAME psychoacoustic model → scalefactors → quantize
notlame:          Audio → Neural network (PsychoNet) → scalefactors → quantize

The neural network learns WHERE to spend bits for best perceptual quality.

Documentation

• How To Train - Simple training guide
• Technical Details - Loss functions, architecture, research
• Improvements - Future enhancements planned

Test Suite

141 tests verify the training pipeline before you train:

make test               # Run all tests
make test-beat-lame     # Verify we beat LAME at all 9 bitrates

Current status: Beats LAME on 8/9 bitrates (STFT), 9/9 bitrates (Mel)

Requirements

• Python 3.8+
• CUDA GPU (recommended)
• LAME encoder (apt install lame)

After Training

Copy the model to notlame encoder:

cp models/psycho_v1.onnx ../notlame-lib/models/

License

MIT