Factor analysis with sampling methods for text dependent speaker recognition

Factor analysis is a method for embedding high dimensional data into a lower dimensional factor space. When data are multimodal we use mixtures of factor analyzers (MFA), which assume statistically independent samples. In speaker recognition, samples are not independent because they depend on the speaker in the utterance. In joint factor analysis and i-vectors, the MFA latent factors are tied at different levels. For example, they can be tied for a segment to extract utterance level information. Tied MFA approaches usually present the drawback that computing the exact posterior of the hidden variables (component responsibilities and latent factors) is unfeasible. For JFA, the preferred approximation consists in computing the responsibilities given a speaker independent GMM and they are fixed during the rest of the process. That implies that the estimated responsibilities for a given sample are independent of the rest of the samples of the utterance not taking into account the shared speaker and channel. We present a novel approximation to jointly estimate responsibilities and latent factors based on sampling the latent factor space. This model differs from previous ones in the hidden variables and parameter estimation; and likelihood evaluation. This approach was tested on the RSR2015 database for text-dependent speaker recognition