User Tools

Site Tools



Andy Sarroff

Create an interface that, given a song with a salient lead vocal:

  • Detects the gender of the lead vocal
  • Source separates the lead vocal
  • Swaps the gender according to some model
  • Resynthesizes the audio

The way I see this happening is via Probabilistic Latent Component Analysis using Ron Weiss' code:

  1. Train the frequency basis functions of two models for male/female vocals using a corpus of a cappella recordings.
  2. Detect gender using log probabilities of models.
  3. Source separate vocal frequency components.
  4. Pitch shift source separated vocals with some formant-sensitive algo
  5. Mix signal back in

There will be lots of technical challenges to this project. For one, it's not easy to perform any kind of source separation. But this might be a fun thing to play with and see how far it goes!

gender.txt · Last modified: 2013/09/24 17:23 by hamr