Pre-process input audio to just use vocals and remove background noise

[kurianbenoy]

Is your feature request related to a problem? Please describe.

Most of the ASR models are trained in clean audio data with minimal background choice. One good way to reduce error rate is using vocals

Describe the solution you'd like

Demucs

• https://github.com/facebookresearch/demucs
• https://github.com/xserrat/docker-facebook-demucs

[kurianbenoy]

Tips shared by Mayank, ex-Amazon guy:

[1:42 am, 13/03/2024] Mayank (Ex Amazon) Hasgeek Hackathon US: I found it hosted already 😀https://replicate.com/cjwbw/demucs
[1:49 am, 13/03/2024] Mayank (Ex Amazon) Hasgeek Hackathon US: Spleeter is much faster, but it outputs only wav format, and that can become too large for OpenAI’s whisper api, with 25 mb limit . There’s whisper on replicate as well, but I think OpenAI’s api is faster
[1:49 am, 13/03/2024] Mayank (Ex Amazon) Hasgeek Hackathon US: https://replicate.com/soykertje/spleeter