Local Browser-Based AI: Running Neural Networks for Audio Stem Separation

I Ran a Neural Network in a Browser Tab to Split a Song into Stems

Aral Roca implemented a local audio stem separation pipeline using an ONNX-exported Demucs v4 model running via WebAssembly. This system processes audio files entirely within the browser tab with zero network requests after the initial page load.

Why This Matters

Traditional audio stem separation relies on costly cloud GPU services that require data uploads, compromising privacy and introducing subscription barriers. While local inference in the browser is bounded by single-thread WebAssembly performance and RAM limits—resulting in 3-5 minute processing times for standard tracks—it eliminates data retention risks and provides a free, private alternative to server-side models for producers and forensic analysts.

Key Insights

• Demucs v4 (htdemucs) uses a transformer-CNN hybrid architecture to capture both long-range dependencies and local spectral patterns, 2026.
• The Web Audio API’s decodeAudioData and OfflineAudioContext provide the infrastructure for raw PCM decoding and resampling to 44100 Hz.
• ONNX Runtime Web enables browser-based inference by leveraging WebAssembly to execute model weights once cached, bypassing remote GPUs.
• Client-side separation ensures privacy for unreleased material as audio bytes exist only in device memory and never leave the machine.
• Web Workers are utilized to run inference on a separate thread to prevent blocking the UI during heavy computation.

Practical Applications

• Transcription Aid: Jazz musicians isolate piano stems to catch harmonic details buried in a mix. Pitfall: Low-bitrate source files (<320kbps) produce significant artifacts.
• Sample Archaeology: Hip-hop producers extract clean drum breaks from vintage recordings. Pitfall: Dense wall-of-sound arrangements cause instrument bleeding into the ‘other’ stem.
• Accessibility: Hard-of-hearing users boost vocal stems and attenuate instruments for clearer dialogue. Pitfall: Mono recordings fail to provide the spatial cues necessary for effective source distinction.

References:

• https://dev.to/aralroca/i-ran-a-neural-network-in-a-browser-tab-to-split-a-song-into-stems-10mk