This is a simplified guide to an AI model called Demucs-Prod maintained by Ardianfe. If you like these kinds of guides, you should subscribe to the AImodels.fyi newsletter or follow me on Twitter.

Model overview

demucs-prod is a state-of-the-art music source separation model created by Facebook Research and maintained by ardianfe. It is capable of separating drums, bass, and vocals from the rest of the musical accompaniment. demucs-prod is based on a hybrid spectrogram and waveform U-Net architecture, with the innermost layers replaced by a cross-domain Transformer Encoder. This allows the model to effectively leverage both the spectral and temporal domains for improved separation quality.

Similar open-source music separation models include demucs, demucs, and all-in-one-audio. However, demucs-prod stands out with its advanced Hybrid Transformer architecture, which achieves state-of-the-art separation performance.

Model inputs and outputs

Inputs

• Audio: The audio file to be processed, in any format supported by torchaudio.

Outputs

• Drums: The separated drum track.
• Bass: The separated bass track.
• Vocals: The separated vocal track.
• Other: The remaining musical accompaniment.

The output tracks are provided as individual stereo WAV or MP3 files, sampled at 44.1 kHz.

Capabilities

demucs-prod is a highly capable music source separation model that can effectively isolate the drums, bass, and vocals from a musical mix. It leverages a hybrid deep learning architecture to capture both spectral and temporal features, leading to impressive separation quality. The model has been trained on a large dataset of musical tracks, including the MUSDB HQ dataset, and can handle a wide variety of musical genres and styles.

What can I use it for?

demucs-prod can be a valuable tool for a variety of music-related applications and projects. For example, it can be used to create "stem" versions of songs, where the individual instrument and vocal tracks are separated and can be processed or remixed independently. This can be useful for music producers, DJs, and audio engineers who need to work with the individual components of a song.

Additionally, the separated tracks can be used for karaoke or music education applications, where the vocals or other specific instruments can be isolated and highlighted. The model can also be used for audio restoration and cleanup, where the separated tracks can be used to reduce unwanted elements or artifacts in the original mix.

Things to try

One interesting aspect of demucs-prod is its ability to handle a variety of input formats and provide flexible output options. Users can experiment with different input audio formats, such as WAV, MP3, or FLAC, and choose to output the separated tracks as either WAV or MP3 files. Additionally, the model supports options for adjusting the segment length, number of parallel jobs, and clip mode to optimize performance and quality for different use cases.

Another area to explore is the model's ability to separate more than just the drums, bass, and vocals. The demucs-prod model also includes an experimental 6-source version that adds "guitar" and "piano" as additional separation targets, although the quality of the piano separation is currently limited.

If you enjoyed this guide, consider subscribing to the AImodels.fyi newsletter or following me on Twitter for more AI and machine learning content.