Skip to Main Content
As a first step in developing an emotion recognition system from human voice, it is necessary to collect relevant set of emotionally rich utterances that will be used for system training. Thus, a first emotional speech corpus of Croatian language (KEG) was built and annotated. The collection and annotation process together with some interesting statistical properties of the designed corpus are described in this paper. Utterances were collected from both male and female speakers, from child age to adults, verbally expressing their emotions. Materials were taken from Internet and other public media sources, with the total duration of approximately 40 minutes. Emotion classification used for annotation has been based on 5 discrete emotional states: happiness, sadness, fear, anger and neutral state. For each of the non-neutral emotional states, the perceived intensity was also annotated in 10 steps. Preliminary KEG evaluation was performed by building and testing an emotion recognition system based on this specific corpus. Initial results are presented in this paper.