Before you begin
  • Compressed file size and working memory are different.
  • Check the download dialog before starting a new model.
  • Export saved results before clearing site data.

A small file can become a large working buffer

Audio is decoded into samples before separation. A three-minute stereo buffer at 44.1 kHz using 32-bit floating-point samples contains 180 × 44,100 × 2 × 4 = 63,504,000 bytes, about 60.6 MiB. Four such output stems alone use about 242 MiB. The input, model, GPU work, display and export buffers require additional memory.

This explains why a compact MP3 is not a reliable prediction of processing cost. Stemix limits an input file to 256 MB and its audio duration to 10 minutes, but those limits do not guarantee every accepted file will fit every device.

Separate network downloads from your audio

The model, inference runtime and audio codec components are downloaded when needed and stored in the browser’s component cache. Only the chosen 2-, 4- or 5-stem model is requested. The confirmation dialog shows what is missing and its total size. Cached resources can be reused, but a changed version or cleared cache needs another download.

Your selected audio and local history are stored separately from those components. Clearing all site data can remove both the processing cache and your saved results. Export anything important first; local history is not a backup service.

Try these checks in order

  1. Test a short, known-good WAV or MP3 instead of retrying a full batch.
  2. Use a current Chrome or Edge with hardware acceleration and WebGPU available.
  3. Close unrelated memory-heavy tabs and retry the short file.
  4. If only one format fails, try another authorized copy. Stemix tries browser decoding first, then its local FFmpeg fallback.
  5. If a download fails, check the connection and try again before clearing the entire site.

Use cancel rather than leaving a stalled task running

The processing view provides a cancel control. Component preparation and local storage also have timeouts so a browser API that never settles should return you to the upload view. If a fresh attempt still fails, note the browser version, operating system, selected stem count, approximate duration and the exact visible error.

Report enough detail without sending the recording

A useful bug report can describe the steps and file format without attaching private audio. The contact page includes the support email and the information to collect. Never include credentials or a private recording unless you have deliberately chosen to share it and have the right to do so. A reproducible short example you own is often more useful than a very large confidential source.

This guide describes the current product controls and implementation. Separation results depend on the recording, device and model; examples are not quality guarantees.

Further reading: Spleeter · Web Audio API · How the tools work