Stemix
Stemix uses pretrained separation models from Deezer Spleeter, converted for browser use and run through ONNX Runtime Web and WebGPU. The upstream model training is not represented as our own. Components and licences are linked below.
Files use browser audio decoding first, with a local FFmpeg WASM fallback when necessary. The processing path converts audio to 44.1 kHz stereo. The model estimates 2, 4 or 5 groups from the mix. Bleed, remaining reverb and lost detail are possible; the original session is not recovered.
| Input limits | 256 MB · 10 minutes |
|---|---|
| Runtime | WebGPU · Current Chrome / Edge with hardware acceleration recommended |
| WAV | 44.1 kHz · stereo · 16-bit PCM |
| MP3 | 320 kbps · local encoding |
| Save all | One Mix per input, packaged in a ZIP |
BPM and key estimates
The analysis uses an audio excerpt rather than a full musical understanding of the song. Short passages, tempo changes, free timing, modes or dense timbres can make estimates unstable. An uncertain result may be shown. These values provide a practice starting point, not evidence of accurate transcription or musicological analysis.
Audio Inspector
The inspector obtains a decoded Web Audio buffer and directly computes sample peak, whole-file RMS, a near-full-scale sample count and a waveform envelope. The current count threshold is absolute sample amplitude of at least 0.999; this is not a count of clipping events in the original recording. Sample rate and channels describe the browser-decoded result and may differ from source metadata.
The inspector has a 64 MB and 10-minute limit and needs neither a separation model nor WebGPU. It does not compute LUFS, reconstruct true peaks or identify complete container metadata. Its JSON report states the measurement scope.
The home-page demo
The demo is an original 8-second synthetic recording made for this site. Simple oscillators, envelopes and deterministic noise create melody, drums and bass, which are summed at the same scale for the full mix. The layers are prepared parts, not AI-separated outputs from a real song. The visual envelopes are derived from the generated sample peaks. You can download the mix to try the tools.
Download the original test loop (WAV, 8 seconds)
Local storage and network requests
Component caches, language preferences and local audio history serve different purposes. Results remain in this browser; clearing site data can remove both results and model caches. Hosting, model delivery, analytics and advertising have their own network activity. Local audio processing does not mean the page never connects to the network. See the privacy policy.
Components and licences
Spleeter · ONNX Runtime Web · FFmpeg WASM · Web Audio API · Full third-party notices