How does Stemverter 3 work?

Stemverter 3 is a macOS front-end GUI for DEMUCS, an open source command line tool developed by Meta labs that processes audio files using stem separation engine powered by Google’s Tensor Flow. According to Meta’s documentation on GitHub, DEMUCS is:

“a state-of-the-art music source separation model, currently capable of separating drums, bass, and vocals from the rest of the accompaniment. Demucs is based on a U-Net convolutional architecture inspired by Wave-U-Net. The v4 version features Hybrid Transformer Demucs, a hybrid spectrogram/waveform separation model using Transformers. It is based on Hybrid Demucs (also provided in this repo) with the innermost layers are replaced by a cross-domain Transformer Encoder. This Transformer uses self-attention within each domain, and cross-attention across domains. The model achieves a SDR of 9.00 dB on the MUSDB HQ test set. Moreover, when using sparse attention kernels to extend its receptive field and per source fine-tuning, we achieve state-of-the-art 9.20 dB of SDR.”

By bringing DEMUCS into a drag-and-drop interface for macOS users on systems 10.13 High Sierra and newer, cflo inc. has provided an accessible solution for users who want to convert their audio files into stems for use in live performance or audio production projects.

Does Stemverter 3 do anything else?


Along with the DEMUCS engine, Stemverter 3 also includes many additional features that streamline the conversion process and make file management a breeze, including direct to mp3 conversion, ID3 cloning, batch conversion, and automatic smart file naming. Stemverter 3 is also the only Desktop App that converts files using 100% local processing power as well as writes files to the user’s filesystem for later use. By bringing Stemverter 3 to the market, cflo inc. has made this amazing tool accessible to users who otherwise would not have the access or technical know-how to install the command-line version.

Please note that some songs output will contain transients and other frequencies that may sound filtered in a way that’s not to the user’s ideal spec. This is normal, and a result of the models Stemverter 3 employs that are the reference for which frequencies are filtered during the stem separation process. We will issue updates to Stemverter 3 in order to add new features and fixes, so please be sure to check for updates!