StemSplitter.app is a free stem splitter that runs entirely in your web browser. It separates a song into four stems (vocals, drums, bass and other) using an open AI model, and lets you listen, mix and download the results. We built StemSplitter.app because nearly every stem separation site we tried uploaded our songs to a server, made us wait in a queue and then asked for money once a few free minutes were used up. Modern browsers can run neural networks on the graphics card, so none of that is necessary any more.
What makes StemSplitter.app different
- Local processing. Your audio is decoded and separated on your own computer. It is never uploaded to us or to anyone else.
- No account and no limits. There are no credits, no sign-up, no watermark and no cap on the number of songs.
- The full model. We run the complete HT-Demucs v4 model, not a reduced version, and verified that the browser output matches the original PyTorch implementation.
- Fast on ordinary hardware. With WebGPU a typical song is separated in well under a minute on a mainstream graphics card.
How StemSplitter.app works
When you drop a file into one of the StemSplitter.app tools, your browser decodes it with its built-in audio decoders and resamples it to 44.1 kHz stereo. A background worker then splits the audio into overlapping segments of about eight seconds. Each segment is converted into a spectrogram and passed, together with the raw waveform, through the neural network. The network predicts four stems for each segment, and the segments are blended back together with smooth crossfades, so there are no audible joins. Finally the stems are turned into WAV files in your browser memory, ready to play and download.
The heavy lifting happens on your graphics card through WebGPU and ONNX Runtime Web, an open-source inference engine maintained by Microsoft. If your browser does not support WebGPU, the same model runs on the CPU through WebAssembly. That is slower, but the results are the same.
The AI model and credits
The separation model is HT-Demucs v4 (Hybrid Transformer Demucs), developed by Simon Rouard, Francisco Massa and Alexandre Défossez at Meta AI Research and published at ICASSP 2023. Meta released the model and its weights under the MIT license. We use an ONNX conversion of the official weights that is also published under the MIT license, and parts of the signal processing code are adapted from Microsoft's MIT-licensed WebNN developer preview. We are grateful to all of these authors. StemSplitter.app is an independent project and is not affiliated with Meta or Microsoft.
Tools on StemSplitter.app
All tools on StemSplitter.app use the same engine. They differ only in which result they show first and which download buttons they offer:
- Stem Splitter: all four stems.
- Vocal Remover and Karaoke Maker: the instrumental.
- Acapella Extractor and Vocal Stems: the isolated vocals.
- Vocal Splitter: vocals and music as two tracks.
- Drum Remover: the song without drums, or drums only.
- Song Splitter: every track of a song for remixing.
How StemSplitter.app is funded
Because processing happens on your device, running StemSplitter.app costs very little: static web pages and nothing else. We plan to cover those costs with advertising. Ads on StemSplitter.app never have access to your audio files, which stay in your browser. Read the privacy policy for details on cookies and analytics.
Please respect copyright
A stem splitter is a powerful tool, but it does not change who owns a song. Separating music for personal practice, study and private experiments is common. Publishing, selling or performing stems made from someone else's recording usually requires permission from the rights holders. Our terms of use explain this in more detail.
Contact StemSplitter.app
Questions, bug reports, ideas for new tools or feedback on separation quality are all welcome. Email us at our contact address. If a song separates badly, tell us the genre and what went wrong (for example "vocal reverb left in the instrumental") so we can see whether a different model would help. We read every message, usually within a few days.