Publication

Abstract : Vocal emotion conversion forms a significant post-processing step in text-to-speech synthesis. The existing models in emotion conversion are data intensive and computationally expensive. This paper puts forth a hybrid emotion conversion framework in Telugu using enhanced global variance GMM for spectral conversion. Fundamental frequency conversion is achieved using multi-layer feed-forward ANN for higher-dimensional CWT decomposed F0. The method performs better than baseline least squares GMM in perceptive quality for all 3 emotions considered viz. anger, fear and happiness. Strength of constrained variance (CV)-GMM is utilized for improving quality of synthesized emotional speech. Listening tests indicated that quality is considerably improved with maximum Comparative Mean Opinion Score (CMOS) of 4.42 (anger) in Telugu dataset.