Exploiting i-vector posterior covariances for short-duration language recognition

i-vector, uncertainty, calibration, stacked bottleneck features, language identification

In this work we have proposed an approach that accounts for the uncertainty in the i-vector extraction process in the framework of generative Gaussian models for language recognition.

Linear models in i-vector space have shown to be an effective solution not only for speaker identification, but also for language recogniton. The i-vector extraction process, however, is affected by several factors, such as noise level, the acoustic content of the utterance and the duration of the spoken segments. These factors influence both the i-vector estimate and its uncertainty, represented by the i-vector posterior covariance matrix. Modeling of i-vector uncertainty with Probabilistic Linear Discriminant Analysis has shown to be effective for short-duration speaker identification. This paper extends the approach to language recognition, analyzing the effects of i-vector covariances on a state-of-the-art Gaussian classifier, and proposes an effective solution for the reduction of the average detection cost (Cavg) for short segments.