What would be the optimal size of UBM and T for Language recogntion using I-vectors? Normally, in Language recogntion, we have limited number of classes, i.e. 40 languages, so will it be ok to use small size UBM and T? i.e.
UBM--> 256/128/64 (or should it be 512 or more)?
T---> 100(or should it be 400--600)?
In general, the optimal size should be determined through comparison of models. You should train your model with different sizes of UBM and different dimensionalities of your TV space and then compare the performance with some test data in a classification task.
In any case, I think LID systems typically use much larger UBMs, rarely smaller than 512 components, often using even >1000 dimensions, and T dimensionality around 400 to 600. You can use previous literature for reference if you want a starting point, for example in Martinez et al (2011) they had 2048 dimensions for the UBM (with data from 54 languages) and they compared the performance of their classifier on 23 target languages using either 400 or 600 T dimensions. Here's the link: https://www.researchgate.net/publication/221487591_Language_Recognition_in_iVectors_Space
Another reference to look at: Dehak et al (2011) - Language Recognition via Ivectors and Dimensionality Reduction. In this paper they used 2048 dim for the UBM and they compared different dimensions for the TV space, and the optimal one was 400 dim.
Hope this answers your question!
Conference Paper Language Recognition in iVectors Space.
- Badges
- Science topic
- Similar topics
- Linguistics
- Philology
- Language