What technology does Ezstems use to separate audio tracks?

Ezstems separates audio tracks using Spleeter, an open-source AI model from Deezer Research. When you upload a file, Ezstems runs it through Spleeter's pre-trained models to split the song into separate stems, then lets you download the results. This applies to any audio file you upload through the site's standard upload flow.

The core technology: Spleeter

Spleeter is a music source separation tool built on pre-trained models. It was introduced by Romain Hennequin, Anis Khlif, Felix Voituret, and Manuel Moussallam, and presented as a late-breaking/demo paper at ISMIR 2019 (the International Society for Music Information Retrieval conference). Ezstems credits this work directly on its About page.

In practice, "source separation" means taking a mixed recording and estimating which parts belong to which instrument or voice group. Spleeter does this with neural networks trained on audio, rather than by hand-crafted rules.

Why "pre-trained" matters

Because the models are already trained, you don't need to supply training data or configure anything. You upload a file, and the separation runs against existing models. That is what makes the "stems in seconds" claim on the site possible for a normal user with no machine-learning setup.

What Ezstems actually does with it

Ezstems is the front end; Spleeter is the engine. The site describes its process as:

1. Upload an audio file.
2. Process it using the Spleeter AI model to separate the different audio parts.
3. Download the resulting stems, which start downloading automatically when processing completes.

The site warns "Don't close your browser" during processing and shows a queue position, so separation is handled as a job rather than instantly in the page.

Stem and format options

The upload interface exposes choices that map to how many parts you want and how the output is encoded:

The stem count reflects Spleeter's model variants (for example, a 2-stem split versus a 4- or 5-stem split). Higher mp3 bitrates and WAV give larger, higher-quality files; the cutoff setting trades bandwidth for file size.

Practical implications

• No local install needed. You get Spleeter's separation through a browser upload instead of running the model yourself.
• Quality depends on the model and settings. The number of stems and the cutoff/format you pick affect the result, so choose based on whether you need, say, a quick 2-stem vocal/instrumental split or finer 4–5 part separation.
• Free tier has limits. The site states a 20 uploads/day limit and a 20MB file size cap, with paid subscription options to bypass those limits or get API access. Check the Subscribe page for current terms.

If your goal is simply to understand the mechanism: Ezstems = upload and download layer, Spleeter = the pre-trained AI model doing the actual separation, backed by the ISMIR 2019 Deezer Research work.