Free AI Vocal Remover & Karaoke Maker
Powered by the Demucs AI model running entirely in your browser. Get a clean karaoke track and isolated acapella — nothing uploaded, nothing stored.
Drop an audio file here
MP3, WAV, FLAC, AAC, OGG, M4A — must be stereo
First use downloads the 172 MB AI model · cached for future sessions
How it works
Upload a stereo audio file
Drop any MP3, WAV, FLAC, AAC, or M4A file. The first use downloads the 172 MB Demucs AI model — subsequent sessions load it from your browser's cache in seconds.
Demucs AI separates the stems
The htdemucs neural network model runs locally in your browser via WebGPU or WASM. It separates vocals from drums, bass, and other instruments — typically in 1–3 minutes.
Download both stems
Get the instrumental (karaoke) track and the isolated vocal (acapella) as separate stereo WAV files. Preview both in the browser before downloading.
What is Demucs?
Demucs (Deep Extractor for Music Sources) is an open-source neural network developed by Meta Research for music source separation. The htdemucs variant combines a hybrid Transformer with a U-Net architecture trained on thousands of hours of music. It separates audio into four stems — vocals, drums, bass, and other instruments — and is consistently ranked among the best open-source models on standard music separation benchmarks. This tool runs the full model in your browser via ONNX Runtime — no server, no API key, no account.
Frequently asked questions
What AI model does this use?
This tool runs htdemucs, the hybrid Transformer + U-Net model developed by Meta Research for music source separation. It is widely regarded as the best open-source model for this task. The model runs directly in your browser via onnxruntime-web — no server is ever involved.
Why does the first use take a while?
The Demucs model is 172 MB. On first use it downloads from the CDN and is stored in your browser's IndexedDB. Every subsequent session loads it from disk in a few seconds — the download only happens once per browser.
How long does separation take?
On a desktop with a modern GPU, Chrome uses WebGPU and typically separates a 3-minute song in 1–2 minutes. On Firefox or without a compatible GPU, the WASM fallback runs on your CPU and may take 3–5 minutes. Mobile devices are not recommended — they may run out of memory.
Why must the file be stereo?
The Demucs model was trained on stereo audio and expects two channels as input. A mono file (single channel) cannot be processed — the tool will display an error if your file is mono.
What formats are supported?
Any format your browser can decode: MP3, WAV, FLAC, AAC, OGG, OPUS, and M4A are all supported in modern browsers. The output is always 16-bit stereo WAV for maximum compatibility.
Is my audio file uploaded anywhere?
Never. The AI model runs 100% in your browser. Your audio file never leaves your device — not during the download, not during processing.
Will it work on every song?
Demucs performs well on most commercially mixed recordings. Heavily processed vocals, extreme reverb, or unusual arrangements may produce leakage. The in-page audio preview lets you judge quality before downloading.