Speech Emotion Recognition for Autism Spectrum Disorder using Deep Learning

Abstract

Children who belong to autism spectrum disorder have difficulty identifying emotions and expressing their emotions. Because it is hard to identify the emotions like anger, disgust, fear, happiness, neutral, sad, and surprise in other people and themselves. This can be even more severe when it could not be found at the beginning and may lead to impairment of social communication of the child. Through the proposed systematic methodology child can identify their basic emotions and try to express them. This evolved methodology was developed using python language. For emotion recognition used a deep machine learning model like Recurrent Neural Network (RNN) using Keras with a TensorFlow backend. RNN consists of four layers with two long short-term memory (LSTM) layers. To optimize the performance of the model used Adam optimizer. For the training and testing of the model used online available data. For the classification of the emotion’s valuable features of the audio signal like Zero Crossing Rate (ZCR), Chroma STFT, Mel- Frequency Cepstral Coefficient (MFCC), Root Mean Square (RMS) value, and Mel spectrogram are extracted using the python libROSA library. Due to the lack of the data amount and GPU requirements model’s performance can be decreased. This model performed well with the TESS data corpus with 91% test accuracy.