Neighborhood-adapted GMM for speaker recognition.
Delphine Charlet · Odyssey · 2004
In previous work [1], it was investigated how the neighborhood can be used to estimate a better model for a speaker when few training data is avalaible. In this paper, this work is completed by investigating another way to merge models from the neighbors and by introducing a weight on the neighbor models to be merged. Experiments on a telephone speech database show that using the neighborhood-merged model to initialize the training phase provides improvement compared to the UBM approach, when few training data is available.