Emotion plays a significant role in human perception and decision making whereas, prosodic features plays a crucial role in recognizing the emotion from speech utterance. This paper introduces the speech emotion corpus recorded in the provincial languages of Pakistan: Urdu, Balochi, Pashto Sindhi and Punjabi having four different emotions (Anger, Happiness, Neutral and Sad). The objective of this paper is to analyze the impact of prosodic feature (pitch) on learning classifiers (adaboostM1, classification via regression, decision stump, J48) in comparison with other prosodic features (intensity and formant) in term of classification accuracy using speech emotion corpus recorded in the provincial languages of Pakistan. Experimental framework evaluated four different classifiers with the possible combinations of prosodic features with and without pitch. An experimental study shows that the prosodic feature (pitch) plays a vital role in providing the significant classification accuracy as compared to prosodic features excluding pitch. The classification accuracy for formant and intensity either individually or with any combination excluding pitch are found to be approximately 20%. Whereas, pitch gives classification accuracy of around 40%.

Statistical learning theory has been introduced in the field of machine learning since last three decades. In speech recognition application, SLT combines generalization function and empirical risk in single margin based objective function for optimization. This paper incorporated separation (misclassification) measures conforming to conventional discriminative training criterion in loss function definition of margin based method to derive the mathematical framework for acoustic model parameter estimation and discuss some important issues related to hinge loss function of the derived model to enhance the performance of speech recognition system.