Wave-U-Net

Wave-U-Net was the paper that asked whether separation even needs a spectrogram. Instead of the usual STFT-magnitude pipeline (which throws away phase and depends on front-end hyperparameters), it adapts the U-Net to the raw 1D waveform: repeated downsampling and upsampling blocks compute features at multiple time scales, with an output layer that enforces source additivity and a context-aware prediction framework to suppress boundary artifacts.\n\nOn singing-voice separation it matched the then-state-of-the-art spectrogram U-Net on the same data, and it entered the SiSEC 2018 campaign as submissions STL1/STL2. The best released vocal model (M5-HighSR) self-reports median 4.95 dB vocal / 11.16 dB accompaniment SDR — modest by 2024 standards, but this was the proof of concept that made Demucs and every waveform separator after it thinkable.\n\nThe repo is MIT with pretrained models, and it is still one of the most-cited separation papers. It is thoroughly superseded for production use, but as the historical foundation of end-to-end separation it belongs in any catalog.

• Author-reported (ISMIR 2018, arXiv 1806.03185): comparable to the state-of-the-art spectrogram U-Net on singing-voice separation with the same data\n- Author-reported: M5-HighSR best vocal model, median vocal SDR 4.95 dB / accompaniment 11.16 dB\n- Participated in SiSEC 2018 (STL1/STL2); 2018-era numbers, far below modern separators