A discriminative and robust training algorithm for noisy speech recognition

A combined technique of discriminative and robust training algorithms, referred to as D-REST (discriminative and robust environment-effects suppression training), is proposed for noisy speech recognition. The D-REST technique can separately model the environmental characteristics and phonetic information and thus it can train speech models discriminatively on phonetic variability by eliminating the disturbance of environment-specific effects. According to the experimental results of a Taiwan stock name recognition task over a wireless network, the proposed D-REST algorithm has the potential to improve performance not only on diverse training data but also on noise-type unmatched environments between training and testing. Furthermore, the usage of the D-REST algorithm amounted to a 60% reduction in average word error rate over the performance by the conventional MCE/GPD-based training approach without the environment-effects suppression training technique.