How to evaluate vocal removal: listening, speed and workflow
Compare the result you actually need
A useful instrumental preserves musical cues while reducing distracting singing. A useful acapella preserves the voice while reducing accompaniment. Judge those outputs separately. This page explains an evaluation method; it does not publish a measured accuracy score or certify a particular device.
Use the same source and passages
- Keep an original recording you own or have permission to process. Use the same source file for each comparison.
- Select an exposed verse, a busy chorus and a section where the singer stops. Note the preview start positions.
- Compare Fast AI and Quality AI at similar comfortable listening levels. Do not choose the louder result simply because it sounds more impressive.
- Check the instrumental for readable lyrics, vocal reverb, missing bass notes, softened drum hits and watery cymbals.
- Check the vocal for missing consonants, distorted held notes, instrument bleed and unnaturally changing tone.
- Rehearse or edit with the result. A small vocal trace may be acceptable for practice but unsuitable for a detailed vocal edit.
Record speed without mixing different stages
Note the device, browser, recording duration and processing mode. Measure model download separately from preview and full separation. Repeat loading can benefit from cached model files, so do not compare a cached run with a first-time download as if they were the same task.
For video, include audio extraction and MP4 export as separate stages. Report failures as well as successes. A page-loading score measures website responsiveness, not the time needed to separate a song.
Keep a simple comparison record
Record the source, preview position, method, device/browser, elapsed processing time, vocal residue and instrument changes. Use descriptive listening notes rather than an invented percentage. To publish a demonstration, use the actual output from the disclosed source and processing mode.
A generated practice tone can test playback and file handling. It does not establish separation quality on a human singing performance. Public before-and-after examples need a suitable original or licensed vocal recording.
Report a problem with enough context
Send the tool name, page URL, browser, device and the stage that failed to contact@surstudio.app. Avoid sending private recordings or credentials. When reporting a file-picker issue, distinguish a picker that never opens from a file that is selected but fails to decode.
Use the main vocal-removal guide, processing-mode comparison and mobile guide.