METHOD FOR HYBRID GENERATIVE-DISCRIMINATIVE SEGMENTATION OF SPEAKERS IN AUDIO-FLOW Russian patent published in 2014 - IPC G10L15/00 

FIELD: information technology.

SUBSTANCE: verbal segments are extracted. Acoustic MFCC features of a vector are calculated. Each verbal segment is projected to the space EV of proper voices with a degree of 10 so that a set of Y vectors is obtained. Clustering centres C₁ and C₂ of the Y vectors are determined. Discriminative clustering is performed by calculation of parameters of planes H₁, H₂ and approximate determination of concentration areas of the Y vectors that are homogeneous as to speaker's information. Obtained data on the verbal segments are used for initialisation of VB diarisation based on a variation and Bayesian analysis. Marks of the segments as to the speakers during the whole pronouncing are obtained, on the basis of which correction of clustering centres C₁ and C₂ is performed; with that, operations of discriminative clustering, variation and Bayesian analysis and correction of clustering centres are performed subsequently at several iteration EV-VB stages. At each stage of iterations there performed is an analysis of complete segmentation as to the speakers, and at the absence of variations in segmentation on iteration it is stopped; after that, final segmentation representing the table correspondence between the verbal segments of an input signal and the speaker's index is obtained by Viterbi resegmentation.

EFFECT: improving accurate detection of a speaker for a dialogue in a telephone channel.

4 dwg, 1 tbl