Entropy and Dynamism Criteria for Speech and Audio Classification Applications

Abstract

We describe the audio classification system that uses entropy and dynamism criteria as discrimination features. The main idea of this approach is that the input neural net is considered as a informational channel. Channel tuned to the certain type of information transmits it best of all according to the informational criterion. In our case a multilayer perceptron (MLP) emitted posterior probabilities for speech recognition was used as such information channel. Then two features entropy and dynamism were computed using these posterior probabilities. And finally HMM was used as a classifier. Different experiments demonstrated efficient usage possibilities of entropy and dynamism criteria not only in speech music discrimination tasks but also in other application of audio classification.