Kannada Speakers Speech Emotion Recognition Using Deep Learning Technique: Anger, Fear, Sad and Neutral state

Abstract

In this paper, we have experimenting to identifying the Kannada emotional speech signals recognition using deep learning technique. Speech Emotion Recognition (SER) is a hot research topic in the field of HCI (Human Computer Interaction). To experiment we have created our own Kannada Emotional Speech corpus recording in different android mobile phone, with sampling rate 48 kbps and mono channel and speakers are from the Karnataka and Tamil Nadu boarder region, whose accents is a mixer of Kannada, Tamil and Telugu languages. The emotions are anger, fear, sad and neutral is considered. Totally we have 440 Kannada emotional speech signals. To use this speech signal data in our experiment, we have first down sampled all speech signals into 16 kbps, mono channel and stored as wav files. Then we pass the each sample signals into the pre-emphasize phase, then framing and windowing is done, then we extracted features from each voice frame signal like MFCC, Chroma and ‘Mel’ and then passed to CNN (Convolution Neural Network) to train and testing the voice signals. The CNN is capable to handle small amount of data for training, and with smaller number of parameters. The average Kannada emotion recognition accuracy rate of 93.77% is achieved. All computations are done using Python programming language and used Python libraries like Librosa, Keras, pyaudio, soundfile, sklearn.