Can this be used as a noise cleansing from conversation audio file? about spleeter HOT 2 CLOSED

I am not affiliated with the project, but FWIW :

I grabbed 1:00 to 3:00 of the audio of this BMXer recording himself on a GoPro camera, riding a bike around Los Angeles and speaking to the camera.

https://www.youtube.com/watch?v=a0vgJ3TeJzA

I then ran it through the 2stem spleeter :

spleeter separate -o ./audio_output/ -p spleeter:2stems -i input/00-YOUTUBE/a0vgJ3TeJzA.flac

I then merged the two files into one, with each file panned hard right and left :

ffmpeg -i vocals.wav -i accompaniment.wav -filter_complex "[0:a][1:a]amerge=inputs=2,pan=stereo|c0<c0+c1|c1<c2+c3[a]" -map "[a]" stereo.split.accomp.and.vocals.wav

The result was the bicyclist talking in my left ear, and a bunch of background noise in my right ear. When I removed the right ear, I heard the speaking quite clearly in my left ear. When I removed the left ear, I heard almost no speaking in the right ear. The only time I heard his voice in the right ear was when it was mistakable for a percussive drum sound. I sometimes heard people other than him speaking in the left ear, when he biked past people who were talking.

I believe this is sufficient proof of the concept of using spleeter's 2stem model to isolate spoken human voices from recordings with background noise. Try it yourself and see!

from spleeter.

Hi @AIGyan

We haven't done any sort of evaluation on this task, nor was our model trained on such examples. Speech enhancement and denoising being an active research field, I assume there are more specialized tools to do that out there.

That being said I can only agree with @awesomer feel free to try it out and let us know what you find!

from spleeter.