A lightweight 2D CNN based approach for speaker-independent emotion recognition from speech with new Indian Emotional Speech Corpora

Speech Emotion Recognition (SER) is the process of recognizing emotions by extracting few features of speech signals. It is becoming very popular in Human Computer Interaction (HCI) applications. The challenge is to extract relevant features of speech to recognize emotions with a low computational cost. In this paper, a lightweight Convolutional Neural Network (LCNN) based model has been proposed which extracts useful features automatically. The speech samples are converted into spectrograms of size 224 × 224 for LCNN input. 5 CNN layers and stride are used for down-sampling the feature maps in place of pooling layers which reduces the computational cost. It has been evaluated for accuracy on publicly available benchmark datasets EMOVO (81%), EMODB (87%), and SAVEE (80%). The accuracy of proposed model is also found to be better than SER CNN-assisted model, ResNet-18 and ResNet-34 models. Very few speech datasets are available in Indian ascent. So, authors have created a new Indian Emotional Speech Corpora (IESC) in English language with 600 speech samples recorded from 8 speakers using 2 sentences in 5 emotions. It will be made publicly available for researchers. The accuracy of the proposed LCNN model on IESC is found to be 95% which is better than existing datasets.