INTERSESSION VARIABILITY COMPENSATION FOR AUTOMATIC EXTRACTION OF INFORMATION FROM VOICE

摘要

Disclosed herein is a method for compensating intersession variability for automatic extraction of information from an input voice signal representing an utterance of a speaker, comprising: processing the input voice signal to provide feature vectors each formed by acoustic features extracted from the input voice signal at a time frame; computing an intersession variability compensation feature vector; and computing compensated feature vectors based on the extracted feature vectors and the intersession variability compensation feature vector; wherein computing an intersession variability compensation feature vector includes: creating a Universal Background Model (UBM) based on a training voice database, the Universal Background Model (UBM) including a number of Gaussians and probabilistically modeling an acoustic model space, creating a voice recording database related to different speakers and containing, for each speaker, a number of voice recordings acquired under different conditions; computing an intersession variability subspace matrix (U) based on the voice recording database, the intersession variability subspace matrix (U) defining a transformation from an acoustic model space to an intersession variability subspace representing intersession variability for all the speakers; computing an intersession factor vector (x<SUB>i</SUB>) based on the intersession variability subspace matrix (U), the intersession factor vector representing the intersession variability of the input speech signal in the intersession variability subspace; and computing the intersession variability compensation feature vector based on the intersession variability subspace matrix (U), the intersession factor vector (x<SUB>i</SUB>) and the Universal Background Model (UBM).