Emotion Recognition 97 Development Set Utterances Percentage

Studies of expressive speech have shown that discrete emotions such as anger, fear, joy, and sadness can be accurately communicated, also cross-culturally, and that each emotion is associated with reasonably specific acoustic characteristics [8]. However, most previous research has been conducted on acted emotions. These certainly have something in common with naturally occurring emotions but may also be more intense and prototypical than authentic, everyday expressions [6, 13]. Authentic emotions are, on the other hand, often a combination of different affective states and occur rather infrequently in everyday life. They are, moreover, often restricted to only a few emotions, such as joy or frustration [3, 5, 7]. In the CHIL project, we are interested in acoustic emotion recognition in two of the given scenarios: the Socially Supportive Workspaces and the Connector agent. In the first one, we want to monitor people attending a lecture to give the speaker feedback on the attentive states of the audience; are they positive and laughing, ignorant and bored, or negative and irri-tated? In the Connector scenario, in which somebody tries to reach a person on the phone via the Connector agent, it is of interest to know whether the caller is starting to show some frustration. Since there was no relevant CHIL material recorded and available for our research, we decided to use two other corpora that we considered to be rather close to what we needed. They are recorded in similar circumstances and environments as the CHIL scenarios. For the Socially Supportive Workspaces, we used a corpus of small-group interaction collected at CMU and known as the ISL Meeting Corpus [2]. For the Connector agent, we used a database from the Swedish telephone service company Voice Provider that contains recordings of people interacting with automatic voice response centers [12]. The emotional context of the ISL Corpus is human-human interaction and the emotions conveyed are mainly positive, often associated with laughter. The emotional context of the Voice Provider Corpus is human-machine interaction. The emotions are rather rare and mostly negative due to frustration with the performance of the automatic voice response system. Thus, both databases contain authentic and not acted emotions in settings similar to the CHIL scenarios we want to explore. However, their diverse contexts make the task for the emotion recognizer quite different. In the following, we will first describe