Emotion conversion using Feedforward Neural Networks

An emotion is made of several components such as physiological changes in the body, subjective feelings, and expressive behaviours. These changes in speech signal are mainly observed in prosody parameters such as pitch, duration and energy. In this work, prosody parameters are modified using instants of significant excitation (epochs) and these instants are detected using Zero Frequency Filtering (ZFF) based method. Epoch locations in the voiced speech corresponds to glottal closure instances, and in the unvoiced region it corresponds to some random instants of significant excitation. Prosody parameters for target emotions are predicted from Hindi emotional speech database. In this work, anger and sad emotions are considered as target emotions in the proposed emotion conversion framework. Feedforward neural network models are explored to predict the prosody parameters. Predicted Prosody parameters at syllable level are incorporated into neutral speech to produce the desired emotional speech. After incorporating the emotion specific prosody, perceptual quality of the transformed speech is evaluated by listening tests. Keywords-emotion conversion; zero frequency filtering; pitch; duration; glottal closure instance; feedforward neural network