search

facebookresearch / demucs

Code for the paper Hybrid Spectrogram and Waveform Source Separation
MIT License
8.19k stars 1.04k forks source link

"Buzz" sound being added to 'Other' & 'Vocals' #92

Open lucellent opened 4 years ago

lucellent commented 4 years ago

I recently noticed there was a buzzing sound being added to the synths and vocals. I don't know if it's constant but it's for sure at the beginning and end. Maybe where there's no sound. I double checked and this sound doesn't exist in the original song.

I'm attaching an example of an end of a song. You can hear it without amplifying the sound but if you turn up the volume you can definitely hear it. For me it's very noticeable.

https://clyp.it/p0ckuyea

adefossez commented 4 years ago

Buzzing will happen due to the architecture of the network. You can partially remove it using --shifts=5 for instance. It will probably never go away completely though.

Le mer. 4 mars 2020 à 18:31, lucellent notifications@github.com a écrit :

I recently noticed there was a buzzing sound being added to the synths and vocals. I don't know if it's constant but it's for sure at the beginning and end. Maybe where there's no sound. I double checked and this sound doesn't exist in the original song.

I'm attaching an example of an end of a song. You can hear it without amplifying the sound but if you turn up the volume you can definitely hear it. For me it's very noticeable.

https://clyp.it/p0ckuyea

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/facebookresearch/demucs/issues/92?email_source=notifications&email_token=AAPF3PWX6FPPTKFNOKFF2HLRF2F7RA5CNFSM4LBOLKM2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4ISPAU2Q, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAPF3PU6OR4YMKHYU4J53VDRF2F7RANCNFSM4LBOLKMQ .

-- Alexandre Défossez

lucellent commented 4 years ago

I did try with shifts=50 for the sake of curiosity. Don't hear a huge difference in quality, but the buzz is quieter now. Not gone, but not so loud.

Also... the processing took 3 hours.

adefossez commented 4 years ago

If shifts=X, then it will process the audio X times with different random shifts of the input data. Without shifts, the processing time is roughly 1 second for 1 second of audio. So with shifts=50, you almost get a minute of processing time per second of audio... If you have a GPU around you can try using the light or light_extra model with the -n option and run it on GPU (just replace -d cpu by -d cuda). You need to use the light model as the default one will probably run out of memory on GPU.