I have a corpus with 50 speakers. I am a bit confused on how should I divide the corpus to perform a cross-validation strategy.
I am using a GMM-UBM strategy, so I have to divide the corpus in speakers to train the UBM model and speakers to perform tests. The latter also need to be sub-divided in registered speakers and impostors.
Any material on how to do this would be greatly appreciated.
Thank you very much.
I think that the above strategy works well in your case, but for training the UBM model I would use a more bigger corpus like one of the NIST public databases (NIST SRE 2004, 2005, 2008), that way your UBM model will be more descriptive about the general characteristics of the speakers. Then you can follow the above strategy by distributing the 20 speakers for UBM between impostors and registered.
In a previous research I trained the UBM model using 1661 speakers from the public database NIST SRE 2005.
Conference Paper Speaker Verification Using Accumulative Vectors with Support...
- Badges
- Science method