TIME-DEPENDENT CROSS-PROBABILITY MODEL FOR FEATURE VECTOR NORMALIZATION
Luís Buera, Eduardo Lleida, Antonio Miguel, Alfonso Ortega, Óscar Saz · 2006
In previous works, Multi-Environment Model based LInear Normalization, MEMLIN, and Phoneme-Dependent MEMLIN, PD-MEMLIN, were presented and they were proved to be effective to compensate environment mismatch. Both are empirical feature vector normalization techniques which model clean and noisy spaces with Gaussian Mixture Models, GMMs, and the probability of the clean model Gaussian, given the noisy model one and the noisy feature vector (cross-probability model) is a critical point in both algorithms. In the previous works the cross-model probability was approximated as time-independent. However, in this paper, a time-dependent estimation based on GMM is proposed for MEMLIN and PD-MEMLIN. Some experiments with SpeechDat Car database were carried out in order to study the performance of the proposed estimation of the cross-probability model in a real acoustic environment, obtaining important improvements: 78.48% and 76.76% of mean improvement in Word Error Rate, WER, for MEMLIN and PD-MEMLIN, respectively (70.21% and 75.44% if timeindependent cross-probability model is applied).