SplitSonic - Audio Source Separation & Guitar Removal Tool

A comprehensive Python application for audio source separation and selective instrument removal using state-of-the-art deep learning models. I use to this split stems , mostly removing guitar tracks that I can practice over to hone technique.

🏗️ Application Architecture

Core Components

splitsonic/
├── modelsplitter.py       # Main module with advanced separation algorithms
├── app.py                 # Flask web interface and REST API
├── run_web_interface.py   # Launcher for the web interface
├── requirements.txt       # Core Python dependencies
├── app_requirements.txt   # Web interface dependencies
├── templates/             # Web frontend templates
├── README.md              # This documentation
└── WEB_INTERFACE.md       # Web interface documentation

Architecture Overview

graph TD
    A[Audio Input] --> B[Audio Loading]
    B --> C[PyTorch/Demucs Model]
    C --> D[Source Separation]
    D --> E{Processing Type}
    E -->|Basic Separation| F[Individual Stems]
    E -->|Guitar Removal| G[Mixed Output - Guitar]
    E -->|Selective Filtering| H[Spectral Processing]
    E -->|Vocal Removal| I[Instrumental Output]
    F --> J[Audio Files Output]
    G --> J
    H --> J
    I --> J

Loading

Technology Stack

Huge thanks for all the opensource librarirs that make this possible.

• Deep Learning Framework: PyTorch + TorchAudio
• AI Model: Demucs (Hybrid Transformers for Music Source Separation)
• Audio Processing: LibROSA, SciPy, SoundFile
• Core Computing: NumPy
• Audio Format Support: FFmpeg integration

🎵 Features

Primary Functions

1. separate_stems() - Complete audio source separation

   • Separates audio into: drums, bass, vocals, other (guitars/keyboards)
   • Outputs individual stem files
   • Uses Demucs neural network model

2. remove_guitar() - Traditional guitar removal

   • Removes entire "other" stem containing guitars
   • Also removes keyboards and other non-drum/bass/vocal instruments
   • Fast but less selective

3. remove_guitar_keep_keyboards() - Advanced selective removal

   • Uses spectral filtering to target guitar frequencies
   • Preserves keyboards and other harmonic instruments
   • Configurable filter strength (0.0-1.0)

4. remove_vocals() - Vocal removal for instrumentals

   • Creates instrumental versions
   • Preserves all instruments including guitars

Supported Audio Formats

• WAV, MP3, FLAC, M4A, and other FFmpeg-supported formats
• Stereo and mono input support
• Automatic format conversion

🚀 Usage

Basic Usage

from modelsplitter import separate_stems, remove_guitar, remove_vocals

# Separate audio into individual stems
stems = separate_stems("song.mp3", "./output_stems/")

# Remove guitar while keeping other instruments  
remove_guitar("song.mp3", "song_no_guitar.wav")

# Create instrumental version
remove_vocals("song.mp3", "song_instrumental.wav")

# Advanced: Remove guitar but keep keyboards
remove_guitar_keep_keyboards("song.mp3", "song_selective.wav", guitar_filter_strength=0.9)

Command Line Usage

# Process single file with all operations
python modelsplitter.py input_song.mp3

📋 Installation

Prerequisites

• Python 3.8+ (recommended: 3.11)
• CUDA-compatible GPU (optional, for faster processing)

System Requirements

• RAM: 8GB minimum, 16GB recommended for large files
• Storage: 2GB for model files + space for output files
• GPU: Optional but significantly speeds up processing

🔧 Configuration

Model Selection

Default: htdemucs (Hybrid Transformers Demucs)

• High quality separation
• Slower processing
• Alternative: demucs for faster processing

Audio Processing Parameters

• Sample Rate: Automatically detected and preserved
• Channels: Stereo processing (mono files converted to stereo)
• Bit Depth: 32-bit float internal processing, configurable output

🐛 Troubleshooting

Common Issues

1. Out of Memory: Reduce file size or use CPU processing
2. Slow Processing: Enable GPU acceleration or use lighter models
3. Audio Quality: Check input file quality and format

Debug Mode

Enable verbose logging by setting environment variable:

export DEMUCS_VERBOSE=1

📈 Performance

Typical Processing Times

• 3-minute song: 30-60 seconds (GPU) / 2-5 minutes (CPU)
• Full album: 15-30 minutes (GPU) / 1-2 hours (CPU)

Memory Usage

• Typical: 2-4GB RAM
• Large files: Up to 8GB RAM

🙏 Acknowledgments

• Demucs: Facebook Research team for the source separation model
• PyTorch: For the deep learning framework
• LibROSA: For audio processing utilities

📄 License

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