I’m trying to extract speech from some videos which have noises at the same or higher level than the voice of the speaker. For instance listen from linked point in this video. Is it possible to extract the speech with higher quality from above such recordings so those become easy to understand using source separation/cancelation framework or something else?

In my research I've come across to some recently published approaches (but being a novice in ML/signal processing, I'm not sure what exactly would be helpful) DRNN4ASS A Wavenet for speech denoising and RNNoise

  • this question should be moved to dsp.stackexchange.com – audionuma Jul 8 at 13:15

If your audio allows for source separation and you can phase reverse one channel, then you may be able to remove sound that is in both channels - it really only works well with very defined audio position.

Noise cancellation is unlikely to help here - it can reduce periodic noise, or certain forms of interference, but if they are at higher level than the sound your are interested in then you aren't going to get good results.

The usual guidance is to re-record or overdub.

  • I’m able to find frequent occurances of noise-only segments in the audio with duration of couple or so seconds max. Would that be enough for source separation? IF so which tool would you recommend for it? – Varun Om Jul 9 at 4:46
  • No. That's not what source separation means. Have a wee look at the other questions here on this. Typically if you can phase reverse one channel and add them, sound in both that is identical will cancel. Sources that are separate will not. – Rory Alsop Jul 9 at 7:23

Your Answer

 

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.