Spleeter — AI Audio Source Separation by Deezer

Introduction

Spleeter is a fast, GPU-accelerated audio source separation tool released by Deezer Research. It applies pretrained neural network models to split any audio file into isolated stems (vocals, accompaniment, drums, bass, and other instruments), making it useful for remixing, karaoke, music production, and audio analysis.

What Spleeter Does

• Separates audio into 2, 4, or 5 stems using pretrained U-Net models
• Processes audio faster than real-time on GPU hardware
• Accepts common audio formats including MP3, WAV, OGG, and FLAC
• Outputs separated stems as individual audio files in WAV or other formats
• Provides a Python API for integration into custom audio processing pipelines

Architecture Overview

Spleeter is built on TensorFlow and uses U-Net convolutional neural networks trained on the musdb18 dataset. The models operate on Short-Time Fourier Transform (STFT) spectrograms, applying learned masks to isolate each source before inverse-transforming back to the time domain. Pretrained weights are downloaded automatically on first use.

Self-Hosting & Configuration

• Install via pip with optional GPU support through tensorflow-gpu
• Run from the command line with a single command for quick separation tasks
• Use the Docker image for reproducible environments without dependency management
• Configure output codec, bitrate, and sample rate via command-line flags
• Integrate the Python API into batch processing scripts for large audio collections

Key Features

• Three pretrained models: 2stems (vocals/accompaniment), 4stems, and 5stems
• Processes a typical 5-minute song in under 30 seconds on a modern GPU
• No training required — models work out of the box on any music genre
• Supports batch processing of multiple files in a single run
• Actively used in research and cited in hundreds of academic papers

Comparison with Similar Tools

• Demucs — Meta's model with higher quality output; Spleeter is faster and lighter
• Open-Unmix — reference implementation for research; Spleeter is more production-ready
• LALAL.AI — commercial cloud service; Spleeter runs locally with no usage limits
• iZotope RX — professional DAW plugin; Spleeter is free and scriptable
• Whisper — speech recognition, not separation; Spleeter isolates instruments and vocals

FAQ

Q: How good is the separation quality? A: Spleeter produces clean results for most music. Quality depends on the mix complexity and chosen model (5stems yields finer separation than 2stems).

Q: Does it require a GPU? A: No. It runs on CPU but is significantly faster with a CUDA-compatible GPU.

Q: Can I train my own models? A: Yes. Spleeter provides training scripts and documentation for custom model training on your own datasets.

Q: What audio formats are supported? A: Any format supported by FFmpeg, including MP3, WAV, FLAC, OGG, and AAC.

Sources

• https://github.com/deezer/spleeter
• https://research.deezer.com/projects/spleeter.html