Prosodic feature normalization for emotion recognition by using synthesized speech

Emotion recognition from speech signals is one of the most important technologies for natural conversation between humans and robots. Most emotion recognizers extract prosodic features from an input speech in order to use emotion recognition. However, prosodic features changes drastically depending on the uttered text. In order to normalize the differences of prosodic features related to an uttered text, we used a synthesized speech signal. Most speech synthesizers output speech signals with a "neutral" emotion. After extracting prosodic features from an input speech, it is normalized by using prosodic features extracted from the synthesized speech. We propose two types of normalization, frame-level normalization and vector-level normalization. The experimental results showed that the frame-level normalization is effective for two important emotional dimensions. The average normalized difference was decreased by 0.41% (pleasantness) and 1.14% (arousal).