Split vocals and remove noise entirely in your browser.

A free browser-based audio studio for two common jobs: cleaning background noise out of recordings, and separating vocals and instruments out of music. It is built for podcasters polishing a room recording, creators making karaoke or a cappella tracks, voice actors enhancing a take, and musicians experimenting with stems — all without uploading audio to a cloud editor.

Three engines sit behind one interface: a model-free quick stereo extraction, an AI stem separator running the HT-Demucs network on your device, and an RNNoise denoiser — followed by a full non-destructive DSP editor and per-track WAV/MP3 exports.

Three engines, one drop zone

Quick extraction needs no model at all: it uses stereo phase and channel positioning to pull centered vocals out of accompaniment instantly. AI separation is the high-quality path: the HT-Demucs network (StemSplitio/htdemucs-onnx, disclosed with its model card) runs through onnxruntime-web on WebGPU with a WASM fallback, and mono files are up-mixed automatically. Denoise runs the RNNoise recurrent network in real time via WASM, with a strength slider that blends processed and dry signal. You pick the engine explicitly — there is no silent swapping of methods under vague wording.

AI stem separation in detail

Audio is decoded to 44.1 kHz and processed in roughly 7.8-second segments with weighted overlap-add, which removes the seams you would otherwise hear between chunks. The network returns four stems — drums, bass, other and vocals — and the instrumental is built by combining the non-vocal stems. Per-chunk progress, cancellation and the same WebGPU→WASM safety net as the other tools are built in. Free plans expose the two-track workflow; Pro unlocks all four stems.

Denoising and DSP editing

After processing, the editor offers a full chain of pure, non-mutating operations on a waveform selection: trim, gain in decibels, peak normalization, fade in and out, silence removal, high-pass filtering, gentle compression, and a one-click voice-enhance preset combining the above. An A/B comparison lets you hear processed against original, every action enters undo/redo, and long selections are made by dragging on the canvas or nudging with the keyboard.

Exports and karaoke

Every stem can be downloaded as WAV or, where the browser supports it, MP3; Pro can take the full set as a ZIP. The karaoke export renders the instrumental mix for sing-along use, and nothing is watermarked. As with the other audio tools, formats are probed up front so a format never fails only after a long processing run.

Free and Pro limits

The free tier denoises files up to 30 minutes, separates up to 10 minutes with two exported tracks, and batches five files. Pro raises denoising to 180 minutes, separation to 60 minutes with all four stems, and batches to twenty. No run counts, monthly quotas, or watermarks apply to either tier, and commercial use is included. Quick extraction has no minute gate, though it only yields useful results on stereo, center-mixed material.

Common questions

Does AI separation work on mono or live recordings? Mono is up-mixed automatically, but separation quality reflects the original mix — dense, overlapping material is harder for any source separator.

What is the difference between quick and AI extraction? Quick is instant stereo geometry with no download; AI downloads the HT-Demucs model and produces four real stems with far better bleed control.

Can denoising rescue overlapping speech? Steady background noise is its strength; sudden interferences and other voices are genuinely hard and the strength slider lets you trade artifacts against cleanliness.

Does anything leave my machine? No uploads, no server round-trips; the only network traffic is the one-time model download, and rough takes or unreleased tracks never leave the device.