Spaces:
Running
on
Zero
baicai1145 Apollo-vocal-msst
The open-sourced content includes models for inference at https://github.com/ZFTurbo/Music-Source-Separation-Training and the original weights with fewer training steps. The training was conducted using sucial's project at https://github.com/SUC-DriverOld/Apollo-Training, with a 92-hour high-quality vocal dataset trained for 1 million steps.
https://huggingface.co/baicai1145/Apollo-vocal-msst/blob/main/config_apollo_0129_-21.98.yaml
https://huggingface.co/baicai1145/Apollo-vocal-msst/blob/main/epoch%3D54-val_loss%3D-17.6221.ckpt
https://huggingface.co/baicai1145/Apollo-vocal-msst/blob/main/model_apollo_0129_-21.98.ckpt
Could you please explain what is it and what do you propose?
The model is likely to be more robust to improve voices. The previous model weighed less than 60MB approximately 6 months ago, this one weighs 600MB, updated a month ago.
Interesting, thanks! I tried a 192 Kbit encoded file of a crappy MP3. Spectrum analysis shows the well known LPF around 16 Khz is now filled with audiodata. De rest of the waveform doesn't show much difference.
Do I hear the difference? Not really, but that could be because the quality of the track wasn't all that. There's probably some improvement to make, and still in a scientific phase.
If this would be a real product, it's an idea to keep the original name, and maybe have a lossless FLAC format option, because it could keep the tracks' metadata, while .wav doesn't.
