Development of a transformation algorithm for emotional speech signal using DWT and Adaptive Filter for a Voice Culture Training System

This paper develops an algorithm “Discrete Wavelet Transform with Adaptive Filter” (DWTAF) to transform Neutral speech into emotional speech like Angry, Happy or Sad and this is compared with two other emotion transformation algorithms. The other two algorithms are “Speech Transformation using Statistical Parameters and Pitch Contours” (STSPPC) and “Speech Transformation using Mel Frequency Cepstral Coefficients (MFCC) and Gaussian Mixture Model (GMM)” (STMG). STSPPC calculates statistical parameters of speech like mean and variance and they are used for transformation of emotional speech. STMG performs DTW (Dynamic Time Warping), extracts MFCC and applies a GMM for speech transformation. The proposed algorithm DWTAF, presents a novel approach to model the speech using Discrete Wavelet Transform (DWT) and emotional speech is generated by filtering the speech using an adaptive filter with Least Mean Square (LMS) algorithm. For this purpose a database of 400 English language sentences has been created with 10 sentences uttered by 10 female speakers with 4 emotions each and a comparison between the three transformation algorithms is carried out using Subjective and Objective evaluation tests. The parameters used for Objective evaluation of transformed speech are Segmental SNR (Seg SNR), Log-Likelihood Ratio (LLR) and Weighted Spectral Slope (WSS).