Psychological Motivated Multi-Stage Emotion Classification Exploiting Voice Quality Features

Paralinguistic properties play a more and more decisive role in recent speech processing systems like automatic speech recognition (ASR) or natural text-to-speech systems (TTS). Besides the linguistic information, the so called paralinguistic properties can help solving ambiguities in man-machine-interaction. Nowadays, such man-machine-interfaces can be found for example in call-centers, in driver assistance systems, or at the personal computer at home. There are many different applications for the recognition of paralinguistic properties, e.g. gender, age, voice quality, emotion, or alcohol consumption. Among these properties, the emotional state of a speaker has a superior position because it strongly affects the acoustic signal produced by the speaker in all kind of conversational speech. Emotion recognition has its applications in various fields e.g. in call centers to detect angry customers, in entertainment electronics, in linguistics, and even in politics to analyse speeches of politicians to train the candidates for election campaigns. Various attempts show quite good results in the case of speaker dependent classification (Lugger & Yang, 2006; McGilloway et al., 2000; Nogueiras et al., 2001). But the hardest task and also the most relevant in practice is the speaker independent emotion recognition. Speaker independent means that the speaker of the classified utterances is not included in the training database of the system. He is unknown for the classifier and the deduced learning rules in the training phase. Up to now, a good speaker independent emotion recognition could only be achieved by using very large feature sets in combination with very complex classifiers (Schuller et al., 2006; Lee & Narayanan, 2005). In (Schuller et al., 2006), an average recognition rate of 86.7% was achieved for seven emotions by using 4000 features and support vector machine as classifier. In this work, our goal is to further improve the classification performance of the speaker independent emotion recognition and make it more applicable for real-time systems. Therefore, we use the same German database consisting of 6 basic emotions: sadness, boredom, neutral, anxiety, happiness, and anger. For lack of relevance disgust is ignored. But in contrast to other approaches, we focus on the extraction of less but more adapted features for emotion recognition. Because for real-time systems the feature extraction is the most time consuming part in the whole process chain, we try to reduce the number of features. At the same time we study multi-stage classifiers to optimally adjust the reduced feature number to the different class discriminations during classification. In comparison to support vector machines or neural networks, the Bayesian classifier we use can be O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg