Speech Emotion Recognition Using Magnitude and Phase Features

For the majority of people, speaking is a natural method to interact with others, and speech is a tool for understanding others. Speech Utterances are statements made by the Speaker in a discussion that transmit knowledge or a message. Speech processing is the study of how to analyze and process human speech signals or voice samples. The speech utterances include the speaker’s individual qualities, including feelings, bodily functions, and information about the environment. In this paper we have adopted the Constant-Q cepstral-coefficients (CQCC) and relative phase, from the relative phase feature further we derive the Linear-prediction residual—Relative phase (LRP-RP), LPA estimated speech based RP (LPAES-RP). Additionally, to improve system performance, the number of features from the LPAES-RP and the CQCC were integrated. Along with these features deep feed-forward neural networks were used to train the model and softmax layers is used to recognize the given speech emotions. The accuracy level of 92.78% is obtained when the LPAES-RP and CQCC features are combined, out of all the specified systems.