Document Type

Theses, Ph.D


Available under a Creative Commons Attribution Non-Commercial Share Alike 4.0 International Licence

Publication Details

A thesis presented to the Technological University Dublin, Digital Media Centre (DMC) for the degree of Doctor of Philosophy, June 2014.


Knowledge of human perception of emotional speech is imperative for the development of emotion in speech recognition systems and emotional speech synthesis. Owing to the fact that there is a growing trend towards research on spontaneous, real-life data, the aim of the present thesis is to examine human perception of emotion in naturalistic speech. Although there are many available emotional speech corpora, most contain simulated expressions. Therefore, there remains a compelling need to obtain naturalistic speech corpora that are appropriate and freely available for research. In that regard, our initial aim was to acquire suitable naturalistic material and examine its emotional content based on listener perceptions. A web-based listening tool was developed to accumulate ratings based on large-scale listening groups. The emotional content present in the speech material was demonstrated by performing perception tests on conveyed levels of Activation and Evaluation. As a result, labels were determined that signified the emotional content, and thus contribute to the construction of a naturalistic emotional speech corpus. In line with the literature, the ratings obtained from the perception tests suggested that Evaluation (or hedonic valence) is not identified as reliably as Activation is. Emotional valence can be conveyed through both semantic and prosodic information, for which the meaning of one may serve to facilitate, modify, or conflict with the meaning of the other—particularly with naturalistic speech. The subsequent experiments aimed to investigate this concept by comparing ratings from perception tests of non-verbal speech with verbal speech. The method used to render non-verbal speech was low-pass filtering, and for this, suitable filtering conditions were determined by carrying out preliminary perception tests. The results suggested that nonverbal naturalistic speech provides sufficiently discernible levels of Activation and Evaluation. It appears that the perception of Activation and Evaluation is affected by low-pass filtering, but that the effect is relatively small. Moreover, the results suggest that there is a similar trend in agreement levels between verbal and non-verbal speech. To date it still remains difficult to determine unique acoustical patterns for hedonic valence of emotion, which may be due to inadequate labels or the incorrect selection of acoustic parameters. This study has implications for the labelling of emotional speech data and the determination of salient acoustic correlates of emotion.