NOISE COMPENSATION METHOD USING SIMULTANEOUS ESTIMATION OF EIGEN ENVIRONMENT AND BIAS COMPENSATION VECTOR

摘要

A noise compensation method using simultaneous estimation of eigen environment and a bias compensation vector is provided to implement a commonly applicable noise compensating method which is not affected by difference between training environment and speech recognition environment. R noise speech databases are organized and a feature vector is extracted from each of the noise speech databases. R noise speech models comprised of M Gausian mixtures are obtained. One original speech model comprised of the M Gausian mixtures is obtained. A compensation vector set comprised of M compensation vectors is obtained, wherein the M compensation vectors indicate difference between the M Gausian mixtures of the noise speech models and those of the original speech model. R compensation vector sets corresponding to the R noise speech models are obtained. The M compensation vectors are connected to each of the R compensation vector sets and thereby R super vectors are formed. K eigen vectors are obtained from the R super vectors. A basis vector for compensating a bias between training environment and speech recognition environment is obtained.