I have a data set of 50 speakers, with 20 utterances of each speaker. In order to train a GMMUBM system can I use features from the first 10 utterances of each speaker to train the UBM model, and next 5 utterances for adapting user-specific models and features from the next 5 utterances to test the system.?
According to your data set, you have the 50X20=1000 number of utterances.
Note UBM is a big size soft presentation of GMM.
To train the UBM, it will be batter to consider 1024 Xn size utterances, whree n=1,2,3....
Note: In the utterance, there should be well balanced of the meal andfemale utterances is required, otherwise during the testing, the female utterance will identify as male.
If you used same person utterance, then the few acoustic fertures are dominated in UBM, that makes the result baised.
Some of relevant papers towards the problem, I have included in research gate or you can get from ACM digital library or Springer digital library.
Thanks
@Muaaz as per MSR Identity Toolkit, there are 630 speakers (TIMIT), and for background training toolbox take 530 speakers (192 female/ and 438 male) and rest of 100 for testing(30 female and 70 male). The UBM is also the speakers other than your target speakers. About Gender balancing, as per my experiments, you don't need to have balance dataset. The optimal UBM size can be 512, about that, I havn't seen any much improvement rather than processing cost.
UBM is created for ( One time). If the initial size of the data set is well enough to create the UBM, it will be enhanced the recognition procedure that effect in testing module.
- Badges
- Science topic