Discriminative codebook design for critical word recognition using evolution strategies

This paper introduces a new approach to vector quantizer (VQ) codebook design for speech recognition systems. Within a discriminative training procedure the codevectors are adjusted with the goal of direct minimization of the recognition error rate of a given recognizer according to the principle of minimum classification error (MCE). Because of the complicated optimization surface and the real valued vector elements evolution strategies (ES) are used for successive modifications of the codevectors. The ES-based method can be applied to any existing VQ-based recognition system. The paper concentrates on the application of ES-optimized codebooks for efficient critical word class expert recognizers. In experiments the performance improvement was examined by including ESVQ-recognizer experts into a 20-word speaker-independent German digit and command word recognition system. The recognition error rate could be reduced to 1.95% with a discrete HMM-recognizer compared to 3.35% without ESVQ-experts