PercGAN

PercGAN: A computationally inexpensive approach to high fidelity stereo audio sample generation.

Pretrained Models

Kick Drum

Kick drum generation model trained on curated kick drum samples.

Kick vs. Generated Comparison:

Snare Drum

Snare drum generation model, focused on producing punchy, tight snare sounds.

Snare vs. Generated Comparison:

Model Usage

1. Prerequisites

• Optional but highly recommended: Set up a Python virtual environment.
  • Audio loader package librosa requires an outdated version of Numpy
• Install requirements by running pip3 install -r requirements.txt

2. Generate audio from pretrained models

Use the generation script with command line arguments:

# Generate 2 kick drum samples with default output path
python src/generate.py --type kick --count 2

# Generate 5 snare samples with custom output path
python src/generate.py --type snare --count 5 --output_path my_samples

Parameters:

• --type: Type of audio to generate (kick or snare)
• --count: Number of samples to generate (integer value)
• --output_path: Directory to save generated audio files (default: "outputs")

Technical Approach

PercGAN combines these audio generation techniques:

• Mel-Spectrogram Representation: For efficient learning of frequency patterns
• Progressive Growing: Training on increasingly higher resolution spectrograms
• Style-Based Generation: Using StyleGAN for better style control
• Multi-Scale Spectral Losses: Specialized frequency, decay, and coherence losses
• Griffin-Lim Reconstruction: Converting spectrograms back to audio

Custom Training

To train on your own audio samples:

1. Collect one-shot audio samples (<1.5 seconds each)
2. Update the model parameters in src/utils/model_params.json
3. Encode your samples with src/data_processing/encode_audio_data.py
4. Train with src/stereo_sample_gan.py
5. Generate new samples with the unified generation script

License

This project is licensed under the MIT License - see the LICENSE file for details.