AutoMix - Neural DJ Transition Generation

Generate professional-quality DJ transitions between tracks using AI. AutoMix uses conditional diffusion models and differentiable audio effects to create smooth, musically-aware transitions.

Features

• One-command interface: Simple CLI for data preparation, training, and inference
• Auto device detection: Works on CUDA GPUs, Apple Silicon (MPS), or CPU
• Multi-GPU training: Built-in DDP support for cluster training
• Differentiable mixing: End-to-end trainable audio effects chain
• Stem-aware: Uses demucs for source separation (drums, bass, vocals, other)

Quick Start

Installation

Requires uv for Python/dependency management:

# Clone the repo
git clone https://github.com/clawd139/automix.git
cd automix

# Install with uv (automatically uses Python 3.12)
uv sync

# Verify installation
uv run automix info

Basic Usage

# Generate a transition between two tracks (no training required)
uv run automix mix track_a.mp3 track_b.mp3 -o transition.wav

# With a trained model
uv run automix mix track_a.mp3 track_b.mp3 --model runs/run1/best.pt -o transition.wav

Training Pipeline

# 1. Prepare training data from your music library
uv run automix prepare --tracks ~/Music/DJ_Library --output processed --max 100

# 2. Train the model
uv run automix train --data processed --output runs/run1 --steps 100000

# 3. Generate transitions with your trained model
uv run automix mix track_a.mp3 track_b.mp3 --model runs/run1/best.pt -o output.wav

Commands

automix info

Show system information and available capabilities:

uv run automix info

automix prepare

Prepare training data from a track library:

uv run automix prepare --tracks ~/Music/DJ_Library --output processed --max 100

Options:

• --tracks, -t: Path to track library directory (required)
• --output, -o: Output directory for processed data (default: processed)
• --max: Maximum number of track pairs to process
• --model: Demucs model to use (default: htdemucs)
• --device: Device for processing (auto, cpu, cuda, mps)

automix train

Train the transition model:

# Single GPU/MPS
uv run automix train --data processed --output runs/run1 --steps 100000

# Multi-GPU (auto-detected)
uv run automix train --data processed --output runs/run1 --steps 100000

# Specify GPU count
uv run automix train --data processed --output runs/run1 --gpus 4

Options:

• --data, -d: Path to processed data directory (required)
• --output, -o: Output directory for checkpoints (default: runs/run1)
• --steps: Maximum training steps (default: 100000)
• --batch-size: Batch size (auto-detected if not specified)
• --lr: Learning rate
• --resume: Resume from checkpoint
• --gpus: Number of GPUs (auto-detected)
• --wandb: Enable Weights & Biases logging
• --synthetic: Use synthetic transition dataset

automix mix

Generate a DJ transition between two tracks:

uv run automix mix track_a.mp3 track_b.mp3 -o transition.wav

Options:

• --output, -o: Output path (default: output/transition.wav)
• --model, -m: Path to trained model checkpoint
• --duration: Transition duration in seconds (default: 30)
• --guidance: Classifier-free guidance scale (default: 2.0)
• --steps: Number of diffusion steps (default: 50)
• --visualize: Generate parameter visualization
• --device: Device for inference (auto, cpu, cuda, mps)

Docker (for GPU clusters)

Build and run on Lambda Cloud or any CUDA machine:

# Build
docker build -t automix .

# Run training
docker run --gpus all -v /path/to/data:/data -v /path/to/output:/output \
    automix train --data /data --output /output --steps 100000

# Run inference
docker run --gpus all -v /path/to/tracks:/tracks -v /path/to/output:/output \
    automix mix /tracks/a.mp3 /tracks/b.mp3 -o /output/transition.wav

Architecture

AutoMix uses a conditional diffusion model architecture:

1. Track Analysis: Extract features (BPM, key, structure, mel spectrograms) from both tracks
2. Stem Separation: Use demucs to separate drums, bass, vocals, other
3. Parameter Generation: Diffusion model predicts effect parameter curves
4. Differentiable Mixing: Apply parameters through differentiable DJ effects chain
5. Audio Output: Generate the final mixed transition

Effect Parameters

The model predicts time-varying parameters for:

• Per-stem: gain, 3-band EQ (low/mid/high), low-pass filter, high-pass filter
• Global: crossfader position, reverb send, delay send

Development

# Install dev dependencies
uv sync --all-extras

# Run tests
uv run pytest

# Format code
uv run black automix/
uv run isort automix/

License

MIT

Citation

@software{automix2024,
  author = {Clawd},
  title = {AutoMix: Neural DJ Transition Generation},
  year = {2024},
  url = {https://github.com/clawd139/automix}
}