Abstract:
The Sinhala language suffers from the lack of appropriate emotional speech database for research into speech emotion recognition. In this paper, we represent our efforts to develop an emotional speech database, as it would facilitate future research on speech emotion recognition in the Sinhala Language. The database is planned to be built with 250 speech corpora of five emotions happiness, anger, neutral, sadness and fear containing 50 speech corpora per each emotion. The input data has been analyzed in three categories, namely training, validation and testing. 100 samples were selected to split for the processes. Target data is a 5 x 100 matrix indicating the state of emotion for the sentences. Then the paper uses an Artificial Neural Network approach to represent a classification analysis and recognition of emotional speech using the short-term and mid-term features of speech signals. The key features include time domain and frequency domain features such as speech rate, energy, entropy of energy, MFCCs, etc. The important key features are used to distinguish the emotions such as happy and sadness or angry and sadness. 35 Features have been extracted from speech signals which are related to the statistics of energy, pitch, etc. The classification results of the trained and validated network are represented in confusion matrices for performance analysis. The results show that the ANN used as a classifier is a feasible technique and the selected features are robust and effective by predicting overall accuracy of 91.65% and 86.5% for Short term and Mid-term proceedings respectively. As well as with compare to the preliminary researches we infer that there is an association between the language and the emotions.
