Bimodal Emotion Recognition using Speech and Physiological Changes

With exponentially evolving technology it is no exaggeration to say that any interface for human-robot interaction (HRI) that disregards human affective states and fails to pertinently react to the states can never inspire a user’s confidence, but they perceive it as cold, untrustworthy, and socially inept. Indeed, there is evidence that HRI is more likely to be accepted by the user if it is sensitive towards the user’s affective states, as expression and understanding of emotions facilitate to complete the mutual sympathy in human communication. To approach the affective human-robot interface, one of the most important prerequisites is a reliable emotion recognition system which guarantees acceptable recognition accuracy, robustness against any artifacts, and adaptability to practical applications. Emotion recognition is an extremely challenging task in several respects. One of the main difficulties is that it is very hard to uniquely correlate signal patterns with a certain emotional state because even it is difficult to define what emotion means in a precise way. Moreover, it is the fact that emotion-relevant signal patterns may widely differ from person to person and from situation to situation. Gathering “ground-truth” dataset is also problematical to build a generalized emotion recognition system. Therefore, a number of assumptions are generally required for engineering approach to emotion recognition. Most research on emotion recognition so far has focused on the analysis of a single modality, such as speech and facial expression (see (Cowie et al., 2001) for a comprehensive overview). Recently some works on emotion recognition by combining multiple modalities are reported, mostly by fusing features extracted from audiovisual modalities such as facial expression and speech. We humans use several modalities jointly to interpret emotional states in human communication, since emotion affects almost all modes, audiovisual (facial expression, voice, gesture, posture, etc.), physiological (respiration, skin temperature etc.), and contextual (goal, preference, environment, social situation, etc.) states. Hence, one can expect higher recognition rates through the integration of multiple modalities for emotion recognition. On the other hand, however, more complex classification and fusion problems arise. In this chapter, we concentrate on the integration of speech signals and physiological measures (biosignals) for emotion recognition based on a short-term observation. Several advantages can be expected when combining biosensor feedback with affective speech. First