A method using linguistic and acoustic features to detect inadequate utterances in medical communication

We have previously proposed two methods using both linguistic and acoustic features separately to detect inadequate utterances in medical communication. However, some inadequate utterances could not be detected because these methods only considered either linguistic or acoustic features, whereas, in general, people use both features to judge an utterance. In this paper, we propose a method using both linguistic and acoustic features. The linguistic features are based on not only word frequency but also sentence and conversation structures. The acoustic features are based on the variances of power and fundamental frequency (F0). A Support Vector Machine (SVM) is used to learn these two types of features compositely. The experimental results showed that the precision of proposed method using both linguistic and acoustic features increased 6% from the traditional recognition method and recall of the proposed method increased 14% from the traditional method.