Speech based emotion recognition in Tamil and Telugu using LPCC and hurst parameters — A comparitive study using KNN and ANN classifiers

Speech based emotion recognition finds numerous applications in automated speech services such as interactive voice recognition systems. It has great implications in investigative application like lie detectors, in medical applications such as diagnosis of mental depression etc. This work aims at developing a feature based emotion recognition system. The speech recordings with the emotions-anger, happiness and sadness in Tamil and Telugu languages are used for this work. Noise and other disturbances from the speech waveforms were separated from the raw speech signals using pre-processing. Two features-Linear Predictive Cepstral Coefficients (LPCC) and Hurst Parameter were extracted. Based on the statistical parameters obtained from these features, the classification was done. Classifiers - K-Nearest Neighbor (KNN) and Artificial Neural Network (ANN) were used to identify respective emotions. The performance of the classifiers were compared in terms of accuracy, precision and recall. The classification system was analyzed in terms of individual features and combination of features. The results showed that when individual features were considered, for both the languages Hurst parameter gave more accuracy, precision and recall when compared to LPCC. Also the combination of these features render a far better performance.