Consonant-Vowel Recognition in Emotional Environment using Trajectory based Stochastic Feature Mapping

The characteristics of consonant-vowel (CV) units differ from one emotion to other emotions. Therefore, the existing CV recognition systems fail to recognize CV units in the emotional environments. The effective way of conveying messages by human beings is by expressing their emotions during conversations effectively. Therefore, in this regard, we propose the CV recognition method in the emotional environments, adaptable to varying emotional moods of the speakers. CV recognition system has been explored to transform from emotional MFCC features to neutral MFCC features. We have proposed method for increasing the accuracy of consonant-vowel (CV) in Indian languages for the emotional speech. In this work, we have developed a mapping method based on MFCC feature transformation framework for developing CV recognition system in the emotion environments. In the proposed method, we are using trajectory based stochastic feature mapping (TSFM) method which is used to map emotional MFCC (Mel Frequency Cepstrum coefficient) features to neutral MFCC features. In the proposed method, we have recognized consonant-vowel in two stages. In the first stage, we have recognized vowels using HMM, while in the second stage, consonants are recognized using Support Vector Machines (SVM). After, normalized performance scores from HMM and SVM are merged for CV recognition, the average performance of (HMM+SVM) is increased by using TSFM from 55.84% to 63.1% for female speaker and from 53.31% to 61.47% for male speaker in three emotions (anger, sadness and neutral) for CV units.