An essential step to achieving human-machine speech communication with the naturalness of communication between humans is developing a machine that is capable of recognising emotions based on speech. This thesis presents research addressing this problem, by making use of acoustic and prosodic information. At a feature level, novel group delay and weighted frequency features are proposed. The group delay features are shown to emphasise information pertaining to formant bandwidths and are shown to be indicative of emotions. The weighted frequency feature, based on the recently introduced empirical mode decomposition, is proposed as a compact representation of the spectral energy distribution and is shown to outperform other estimates of energy distribution. Feature level comparisons suggest that detailed spectral measures are very indicative of emotions while exhibiting greater speaker specificity. Moreover, it is shown that all features are characteristic of the speaker and require some of sort of normalisation prior to use in a multi-speaker situation. A novel technique for normalising speaker-specific variability in features is proposed, which leads to significant improvements in the performances of systems trained and tested on data from different speakers. This technique is also used to investigate the amount of speaker-specific variability in different features. A preliminary study of phonetic variability suggests that phoneme specific traits are not modelled by the emotion models and that speaker variability is a more significant problem in the investigated setup. Finally, a novel approach to emotion modelling that takes into account temporal variations of speech parameters is analysed. An explicit model of the glottal spectrum is incorporated into the framework of the traditional source-filter model, and the parameters of this combined model are used to characterise speech signals. An automatic emotion recognition system that takes into account the shape of the contours of these parameters as they vary with time is shown to outperform a system that models only the parameter distributions. The novel approach is also empirically shown to be on par with human emotion classification performance.

Automatic emotion recognition has long been a focus of Affective Computing. It has become increasingly apparent that awareness of human emotions in Human-Computer Interaction (HCI) is crucial for advancing related technologies, such as dialogue systems. However, performance of current automatic emotion recognition is disappointing compared to human performance. Current research on emotion recognition in spoken dialogue focuses on identifying better feature representations and recognition models from a data-driven point of view. The goal of this thesis is to explore how incorporating prior knowledge of human emotion recognition in the automatic model can improve state-of-the-art performance of automatic emotion recognition in spoken dialogue. Specifically, we study this by proposing knowledge-inspired features representing occurrences of disfluency and non-verbal vocalisation in speech, and by building a multimodal recognition model that combines acoustic and lexical features in a knowledge-inspired hierarchical structure. In our study, emotions are represented with the Arousal, Expectancy, Power, and Valence emotion dimensions. We build unimodal and multimodal emotion recognition models to study the proposed features and modelling approach, and perform emotion recognition on both spontaneous and acted dialogue. Psycholinguistic studies have suggested that DISfluency and Non-verbal Vocalisation (DIS-NV) in dialogue is related to emotions. However, these affective cues in spoken dialogue are overlooked by current automatic emotion recognition research. Thus, we propose features for recognizing emotions in spoken dialogue which describe five types of DIS-NV in utterances, namely filled pause, filler, stutter, laughter, and audible breath. Our experiments show that this small set of features is predictive of emotions. Our DIS-NV features achieve better performance than benchmark acoustic and lexical features for recognizing all emotion dimensions in spontaneous dialogue. Consistent with Psycholinguistic studies, the DIS-NV features are especially predictive of the Expectancy dimension of emotion, which relates to speaker uncertainty. Our study illustrates the relationship between DIS-NVs and emotions in dialogue, which contributes to Psycholinguistic understanding of them as well. Note that our DIS-NV features are based on manual annotations, yet our long-term goal is to apply our emotion recognition model to HCI systems. Thus, we conduct preliminary experiments on automatic detection of DIS-NVs, and on using automatically detected DIS-NV features for emotion recognition. Our results show that DIS-NVs can be automatically detected from speech with stable accuracy, and auto-detected DIS-NV features remain predictive of emotions in spontaneous dialogue. This suggests that our emotion recognition model can be applied to a fully automatic system in the future, and holds the potential to improve the quality of emotional interaction in current HCI systems. To study the robustness of the DIS-NV features, we conduct cross-corpora experiments on both spontaneous and acted dialogue. We identify how dialogue type influences the performance of DIS-NV features and emotion recognition models. DIS-NVs contain additional information beyond acoustic characteristics or lexical contents. Thus, we study the gain of modality fusion for emotion recognition with the DIS-NV features. Previous work combines different feature sets by fusing modalities at the same level using two types of fusion strategies: Feature-Level (FL) fusion, which concatenates feature sets before recognition; and Decision-Level (DL) fusion, which makes the final decision based on outputs of all unimodal models. However, features from different modalities may describe data at different time scales or levels of abstraction. Moreover, Cognitive Science research indicates that when perceiving emotions, humans make use of information from different modalities at different cognitive levels and time steps. Therefore, we propose a HierarchicaL (HL) fusion strategy for multimodal emotion recognition, which incorporates features that describe data at a longer time interval or which are more abstract at higher levels of its knowledge-inspired hierarchy. Compared to FL and DL fusion, HL fusion incorporates both inter- and intra-modality differences. Our experiments show that HL fusion consistently outperforms FL and DL fusion on multimodal emotion recognition in both spontaneous and acted dialogue. The HL model combining our DIS-NV features with benchmark acoustic and lexical features improves current performance of multimodal emotion recognition in spoken dialogue. To study how other emotion-related tasks of spoken dialogue can benefit from the proposed approaches, we apply the DIS-NV features and the HL fusion strategy to recognize movie-induced emotions. Our experiments show that although designed for recognizing emotions in spoken dialogue, DIS-NV features and HL fusion remain effective for recognizing movie-induced emotions. This suggests that other emotion-related tasks can also benefit from the proposed features and model structure.