Treinamento multi-estilo e adaptação de modelos via MAP para reconhecimento de fala em ambientes ruidosos

Abstract

The accuracy of voice recognition systems degrades severely when operating in noisy environments. One of the causes pointed out in the literature for this fact is the acoustic mismatch between the environment in which the training locutions were recorded and the one in which the speech recognition system operates. From the mathematical modeling of the inuence of the environment on the speech signal, two ways of improving the rate of correctness of an automatic speech recognition system in these environments were evaluated: multi-style training, using material corrupted by noise Varied but of low intensity to train the system and adaptation of the acoustic models through the maximum a posteriori method. This combination raised the system hit rate by an average of 23.67%.