Speech bandwidth classification using general acoustic features, modified spectral roll-off and artificial neural network

摘要

In this paper we present research work regarding the speech bandwidth classification using artificial neural network as a classifier. The information about speech bandwidth can be used for the purpose of indexing the multimedia material and also for boosting the automatic speech recognition performance by using separate acoustic models for individual acoustic classes. We proposed also new features for bandwidth classification. The AMSR (Average Modified Spectral Roll-off) feature showed encouraging performance and it is especially appropriate for mobile applications with limited processing capabilities (it needs only simple threshold classifier). The best result of 98.3% was achieved using a combination of bandwidth-specific and general acoustic features.