Professor Urmas Sutrop, University of Tartu
Hille Pajupuu, Institute of the Estonian Language
Professor Jens Allwood, University of Gothenburg
The aim of the thesis was to develop a theoretical base for the Estonian Emotional Speech Corpus and to test the validity of the theoretical starting-points on the Corpus material.
The Corpus is now ready as designed (see http://peeter.eki.ee:5000). The results of the research reveal the importance of detailed planning and of the design elements of the Corpus. The theoretical starting-points of the study are relevant and applicable in real situations. Therefore these results could be taken into consideration in the creation of other emotional speech corpora.
What makes this Corpus unique among the other corpora of its kind is the fact that its sentences have different labels according to whether their emotion is carried just by the sound of the sentence or whether the recognition of their emotion from vocal expression may be influenced by the verbal-semantic content. This classification enables the research of emotions both in speech as well as in writing.
Estonian Emotional Speech Corpus is one of the few freely available documented ones that reviews moderately expressed emotions. The Corpus abandoned acted emotions because of their possible stereotypicality and overactedness. The sentences recorded for the Corpus were read out by a so-called ordinary person, who was not dictated what emotion to use while reading.
The Corpus contains 1,234 Estonian sentences that have passed both reading and listening tests. Test takers identified 908 sentences that expressed anger, joy, sadness, or were neutral.
As the emotions of the sentences contained in the Corpus were determined by listeners, some issues of emotion perception came to the fore: 1) Is sentence emotion identifiable purely from vocal cues, without the speaker being seen? 2) Can age affect the identification of emotion? 3) Is the identification of emotion culturally bound? 4) Does identification depend on the listeners' empathy?
For the first question asking if the emotion of a sentence can be identified from non-acted vocal expression without the speaker being seen, results confirmed the supposition that listeners can recognize the moderate expression of non-acted emotions from the voice of a non-professional reeder. Also, the results support the decision that the emotions of the sentences in the Estonian Emotional Speech Corpus should be determined by Estonian adults aged over 30 who speak Estonian as their native language because they are more likely to have acquired the skills for decoding the culture-specific expression of emotions. Furthermore, the results imply that the understanding of emotions depends on cultural factors and social interactions, including the social norms specific to one culture. The interpretation of emotional messages is therefore learned in the course of social interactions. Research has shown, that in the recognition of emotion from vocal cues, empathy is less important than clinical results would suggest. In conducting emotion studies for speech technological purposes, it is obviously unnecessary to exclude non-empathic people from the testers for the reason that they may not recognize the emotions expressed if their low empathy level is not due to mental or developmental disorders.
The Corpus continues to be developed according to the requirements of new research directions. As the Corpus is publicly available and accessible for free, its data can be used for tackling different research challenges.