Recent advances in PD-MEMLIN for speech recognition in car conditions
Luís Buera, Eduardo Lleida, A. Miguel, Alfonso Ortega · 2005
In a previous work, phoneme-dependent multi-environment models based linear normalization, PD-MEMLIN, was presented and it was proved to be effective to compensate environment mismatch. Since PD-MEMLIN transformations have to be estimated from stereo data corpora, and the computational cost is high, two approaches are proposed: coefficient progressive PD-MEMLIN, CPPD-MEMLIN, and blind PD-MEMLIN. The first one consists on a partial normalization of the feature vector, reducing the computational cost, while blind PD-MEMLIN can be applied over any non stereo data corpora, thus the estimation of the transformation is based on an iterative technique from noisy data and a target clean speech model. Some experiments with SpeechDat car database were carried out in order to study the behavior of the proposed techniques in a real acoustic environment. In the previous work, PD-MEMLIN with stereo data and normalizing 13 MFCC coefficients reached 77.67% of improvement. In this paper, CPPD-MEMLEM with only 4 coefficients obtains an average improvement of 72.40%, and blind PD-MEMLIN obtains an average improvement of 73.96%