Adding dimensional features for emotion recognition on speech

Abstract

Developing accurate emotion recognition systems requires extracting suitable features of these emotions. In this paper, we propose an original approach of parameters extraction based on the strong, theoretical and empirical, correlation between the emotion categories and the dimensional emotions parameters. More precisely, acoustic features and dimensional emotion parameters are combined for better speech emotion characterisation. The procedure consists in developing arousal and valence models by regression on the training data and estimating, by classification, their values in the test data. Hence, when classifying an unknown sample into emotion categories, these estimations could be integrated into the feature vectors. It is noted that the results using this new set of parameters show a significant improvement of the speech emotion recognition performance.

A Spanish Corpus for Talking to the Elderly

“…In this section, we describe and analyse each modality annotation. For more information about the annotation procedure, refer to [10] and [12]. Finally, at the same time as the semantic annotation was carried out, the polarity of the transcribed utterances was also labeled by the same annotators.…”

Perceptual Borderline for Balancing Multi-Class Spontaneous Emotional Data

“…Moreover, this subset is strongly dependent of the task. For instance, a political debate on TV [16] or a podcast interview [17] cannot be expected to show the same human emotions than a human-machine scenario [10], [18], [19]. Indeed, fear is not expected in none of the mentioned corpus.…”

“…This work examines the intrinsic properties of the speech samples to find a suitable borderline between majority and minority emotion classes. In a previous work [19] we have proposed the dimensional model of emotions (Valence-Arousal-Dominance, or VAD) as an additional space of the parameter representation, which has demonstrated to improve the emotion recognition performance. On the other hand, some studies [27] have shown that the oversampling of borderline samples is an effective way to deal with imbalance of data and remains the state of the art.…”

Contrasting the Emotions identified in Spanish TV debates and in Human-Machine Interactions

“…Speech signal includes information about the personal characteristics of the speaker, the content of the message delivered or the language used to code it, among others [1]. The analysis of the speech also allows to estimate, to some extent, the current emotional status of the speaker [2,3,4], even the basal mood, or the probability to be suffering a particular mental disease [5]. However, speech may also be influenced by several other variables, such as the habits of the speaker, his personality, culture or the particular task being performed [6,7] This work is aimed to contrast the similarities and differences for the emotions identified in two very different scenarios: human-to-human interaction on Spanish TV debates and human-machine interaction with a virtual agent in Spanish.…”