You do not always need a full desktop audio suite to extract vocals, make an instrumental, or focus on the drums, bass, guitar, or piano in a song. A browser music-source-separation model can analyze the mix and export estimated parts as separate WAV files.
ToolGarden uses an open-source HT-Demucs ONNX model. Audio decoding, model inference, overlap-add processing, and WAV encoding happen locally in the browser. The first run downloads model assets, but the selected song is not uploaded as a remote separation job.
What is audio stem separation?
A released song is usually a stereo mix in which vocals, drums, bass, and other instruments share the same waveform. It cannot be unpacked like a ZIP file to recover the original studio tracks. A separation model instead estimates how much each source contributes by analyzing spectral patterns, rhythm, timbre, and musical context.
The results can work well for practice, karaoke backing tracks, transcription, arrangement study, and mix references, but they are not identical to the original multitrack session. Reverb, distortion, layered performances, and overlapping frequencies can leave audible bleed between stems.
Should you choose four stems or six stems?
| Model | Outputs | Best for | Main tradeoff |
|---|---|---|---|
| Four-stem HT-Demucs | Vocals, drums, bass, other | Vocals, instrumentals, and rhythm practice | More focused vocals and other, without separate guitar or piano |
| Six-stem HT-Demucs | Vocals, drums, bass, other, guitar, piano | Guitar, piano, or six-part analysis | Adds two stems, while other becomes narrower and may sound weaker |
Choose four stems when your goal is vocal isolation or an instrumental. Choose six only when guitar or piano must be available separately. More outputs do not automatically mean better overall quality; they divide the same mixture into smaller categories.
How to split a song in the browser
- Open the stem splitter, then click the upload area or drag a song onto the page.
- Choose the four- or six-stem model. Four stems is the better default for vocals and instrumentals.
- Select only the stems you need. Avoiding unnecessary long outputs reduces result memory use.
- Select “Separate stems” and keep the page open while it decodes audio, downloads and loads the model, separates segments, and encodes WAV files.
- Preview the results, download individual stems, or package every result in a ZIP archive.
Why is the first run slower?
The page is small, but the neural-network models are not. The current four-stem model is about 158 MiB and the six-stem model about 130 MiB. A model is downloaded only when separation begins, verified, and stored in the browser cache so the same browser can usually reuse it later.
After the download, the browser still needs to create an ONNX Runtime session and load the parameters, which can take 10–15 seconds. The smaller files store FP16 weights, but inference still needs more than 1 GB of working memory. Download size and runtime memory are different measurements.
- Use a current desktop version of Chrome or Edge.
- Close memory-heavy tabs and applications before processing.
- Trim a long song to the section you need before separation.
- Do not reload, suspend, or close the page while the model is loading or running.
- Mobile browsers normally have stricter memory limits and are a poor fit for this model.
Does local processing mean no network traffic?
No. The first run needs a connection to download the page code, ONNX Runtime, and model file. The relevant privacy boundary is that the site downloads software and public model assets while the user-selected audio remains in the browser instead of being posted to a remote separation API.
Model, WASM, and Worker requests in the Network panel are expected. To verify a local-processing claim, inspect whether the selected song appears in a request body rather than treating every network request as an upload. Network use normally drops after the required assets are cached.
How should you evaluate the separated audio?
Do not listen to only one isolated track. Recombine vocals, drums, bass, and other to check whether they approximately reconstruct the source, then inspect the stem that matters most. Light metallic artifacts and bleed may be unavoidable; the practical question is whether they interfere with your intended use.
- Karaoke: check for obvious lead-vocal residue and holes left in the instrumental.
- Instrument practice: prioritize stable timing, melody, and continuity over studio-grade isolation.
- Mixing or sampling: audition on headphones and speakers, and check phase, bass, and transient artifacts.
- Transcription or analysis: use the stem with the clearest target voice or instrument.
Common failures and practical fixes
| Symptom | Likely cause | What to try |
|---|---|---|
| Model download fails | The model host is blocked, the connection stopped, or cached bytes are incomplete | Check the connection and retry; clear the cached model if necessary |
| Model appears stuck while loading | The browser is verifying the model and creating a local inference session | Allow 10–15 seconds and avoid repeated clicks or reloads |
| Model cannot be loaded | The browser or device does not have enough available memory | Close other tabs and retry in desktop Chrome or Edge |
| Separation stops partway through | The audio is too long, too many output stems were selected, or inference ran out of memory | Trim the audio, select fewer outputs, and run it again |
| A stem contains audible bleed | Sources overlap heavily in frequency, reverb, distortion, or stereo position | Try the other model and judge the result against the intended use |
Copyright and responsible use
The ability to separate a recording does not automatically grant permission to publish, sell, perform, sample, or redistribute the results. Commercial releases, public videos, performances, and training datasets may involve rights in the recording, composition, and performance. Process audio you own, are authorized to use, or may lawfully use in your jurisdiction.
Key takeaways
The most important stem-separation choice is not “more stems is better.” Use four stems for vocals and instrumentals, and six only when guitar or piano must be isolated. Browser-local inference reduces the privacy exposure of uploading a song, but a large model still requires download time, substantial memory, and realistic expectations about separation quality.