Original examples · Split · public alpha
How to evaluate separated audio stems: bleed and artifacts
Test notes updated October 10, 2026 · by thecreatingco
Judge separated stems by listening to each part, then listening to them together against the original. A stem label does not mean perfect isolation: residual drums in instruments, watery vocal tails and softened attacks can all matter in a remix.
A test pack with known source tracks
We synthesized an original 16-second instrumental at 120 BPM. It has a click-based drum part, a C2 bass tone and C-major keyboard tones. The pack contains the mix and all three independently synthesized source tracks, so you can compare a separator’s output with something known.
Download the instrumental test pack · Download just the mix WAV
There is no voice in this pack. It cannot evaluate vocal extraction or recognition of a human singer. The reference WAVs are source tracks, not Split outputs. Its simple synthesis is also unlike a dense commercial master; use it as a controlled first check, then try your own representative material.
How to evaluate a Split run
- Load the mix, not a reference stem. Wait for model download and processing. Use the current page’s progress messages to distinguish those phases.
- If separation finishes, download the WAVs before leaving. Keep filenames associated with the original mix and the test date.
- Solo drums. Compare the click attacks with
reference-drums.wav. Listen for keyboard or bass leakage between clicks. - Solo bass, then instruments. Compare tone continuity and attacks with their reference files. Check whether notes are missing, pulsing or duplicated in another stem.
- Listen to any vocals output. Since this source contains no voice, audible content there is leakage; silence is not evidence of successful vocal separation on real singing.
- Sum the outputs at their original alignment and unity gain in an editor, avoiding clipping. Compare with the original mix. Do not normalize each stem separately before this check.
Split vs center-channel vocal removal
These approaches solve different problems. Split uses a separation model to estimate vocals, drums, bass and remaining instruments. A center-channel method instead relies on stereo relationships. Audacity’s documented vocal-reduction effect works best under particular conditions and can remove other centered sound; stereo reverb can remain. It is not equivalent to obtaining four isolated stems. Source: Audacity’s vocal reduction and isolation manual.
| Task | Practical choice |
|---|---|
| You have the session’s original multitracks | Use those exports. A model need not reconstruct audio you already possess. |
| You need parts from a mixed recording | Try a separator, then judge bleed and artifacts in the actual remix context. |
| You want to reduce centered singing in stereo | A center-channel technique may help, but check what else is removed. |
When processing appears stuck
First determine whether the page is downloading the model, decoding audio or separating it. A new browser may still need the model. Try a shorter WAV in a desktop browser and keep the original. A file that plays in another application may still fail that browser’s decoder.
Repeatedly reloading can discard progress. If an error is shown, retain its exact text and note the browser, file duration and file type. Download completed results before closing the page; a cached model does not mean your stems are backed up.
Our observed limit
Our earlier hosted Split capture did not finish processing within its test window. That result does not prove Split never works, but it also cannot support a successful-separation demonstration. We therefore supply source references and an evaluation method, with no claimed Split output or speed benchmark. Dense-mix quality and phone performance remain unqualified.
Reuse the examples
These are original learning fixtures. You may reuse and modify them, including commercially. Read the fixture notes and limits. Keep originals and inspect actual exports before delivery.