Demucs
An open-source source separation model released by Meta (Facebook AI Research) in 2019. It uses an encoder-decoder architecture that learns directly from raw waveforms. Later versions introduced a hybrid Transformer structure, making it one of the leading open-source audio source separation tools.
Timeline (last 90 days)
Vocal separation is usually implemented based on spectral masks or deep learning models (such as Demucs and Spleeter).
Stem extraction is a technology that separates mixed audio into independent tracks such as vocals, drums, and bass, with common tools including Demucs.
The latest Hybrid Transformer version of Demucs achieves approximately 9 dB on the Signal-to-Distortion Ratio (SDR).
Demucs, open-sourced by Meta Research, and Spleeter, developed by Deezer, are milestone models in the field of audio source separation, based on U-Net or Transformer architectures.
Demucs was developed by a research team at Meta (Facebook).
Demucs v4 maintains a leading position in multiple international source separation benchmarks, such as the MDX Challenge.
Early versions of Demucs operated directly in the time domain (waveform), while the later Hybrid Demucs (v3) and HTDemucs (v4) adopted a hybrid-domain strategy that processes audio simultaneously in both the time domain and frequency domain.
All Facts (7)
Vocal separation is usually implemented based on spectral masks or deep learning models (such as Demucs and Spleeter).
50%UnverifiedStem extraction is a technology that separates mixed audio into independent tracks such as vocals, drums, and bass, with common tools including Demucs.
50%UnverifiedThe latest Hybrid Transformer version of Demucs achieves approximately 9 dB on the Signal-to-Distortion Ratio (SDR).
50%UnverifiedDemucs, open-sourced by Meta Research, and Spleeter, developed by Deezer, are milestone models in the field of audio source separation, based on U-Net or Transformer architectures.
50%UnverifiedDemucs was developed by a research team at Meta (Facebook).
50%UnverifiedDemucs v4 maintains a leading position in multiple international source separation benchmarks, such as the MDX Challenge.
50%UnverifiedEarly versions of Demucs operated directly in the time domain (waveform), while the later Hybrid Demucs (v3) and HTDemucs (v4) adopted a hybrid-domain strategy that processes audio simultaneously in both the time domain and frequency domain.
50%