Is this a genuine AI model?

Yes. A pretrained SCNet neural source separation model estimates four separate music stems.

Can I process entire tracks on mobile?

No. Browser-based neural inference requires significant memory; this release limits mobile clips to 8 seconds.

PRETRAINED SCNET · FOUR-STEM AUDIO · WAVE 24

AI Vocal Remover & Stem Separation Studio

Separate vocals, drums, bass and other instruments using a real pretrained SCNet neural source-separation model. Preview each isolated track, mix stems and export genuine WAV audio or a complete ZIP.

True AI audio source separationVocals, drums, bass, otherReal WAV stemsAdjustable remix + ZIP

Neural audio separation workspace

Upload a short sample. Trained model processing starts only when requested.

Original sample preview

Listen to your licensed source audio before separating. The model may need a large download on first use and require significant RAM.

Extracted stems

Technical requirements

High-memory AI model: at least 4 GB of device RAM recommended, with modern desktop Chrome/Edge preferred. Mobile can fail due to system memory limits; a low-memory device is explicitly blocked. Individual stems are actual PCM WAV files if inference completes successfully.

Real AI vocal removal, not phase cancellation

This studio loads the pretrained SCNet music source-separation model using @audio/neural-separate (MIT weights). It processes stereo audio at 44.1 kHz and estimates four independent signals: vocals, drums, bass and other. You can preview, download or combine the signals with adjustable mix gains.

Accuracy, download size and device limitations

Source separation is computationally intensive. The neural model and runtime download on demand and may require over a gigabyte of working memory on some devices. Therefore this version intentionally limits clip length and guards low-memory phones. Results are imperfect, especially on dense music, reverberation or overlapping instruments. Nothing is uploaded to Solvory’s server; model assets are requested from their external host.

Related audio tools

Use Audio Trimmer for preparing short clips and Pitch & Tempo Changer for further editing. The Soundboard Studio can play WAV samples.

FAQ

Does this use a trained AI model?

Yes. SCNet ONNX inference separates four stems, rather than using a basic stereo cancellation filter.

Does it work on every phone?

No. Modern high-memory devices may support short clips, but mobile browsers have significant memory limits and can abort inference.

Can it separate a full song?

Not in this browser-first release. Desktop imports are capped at 20 seconds and mobile imports at 8 seconds for safety.

AI vocal remover and four-stem music separation

Real vocal separation is different from simple stereo phase cancellation. This studio uses an SCNet pretrained neural model to estimate vocals, drums, bass and other instrumental material from short stereo clips. You can listen to each generated stem, download PCM WAV, include or exclude stems in a custom remix, adjust gain and batch-download tracks in ZIP format when inference succeeds.

Why song separation is demanding on mobile

Neural music source separation requires large model weights and substantial working memory even for a short sample. The browser edition deliberately restricts desktop and mobile clip lengths. Output accuracy varies by instrument overlap, mix mastering and reverb, and the model cannot perfectly isolate everything. Import only music you have rights to process; use Audio Trimmer to prepare shorter samples, or Soundboard Studio to use exported effects. Model documentation: SCNet music separation.