Speech Emotion Recognition: Features and Classification Models Using CNN

Due to factors such as the considerable influence of cultural background, recording environment, and language employed in emotional utterances, recognising emotion from speech is a tough process. As a result, developing a high-performance emotion identification engine is tough. Even if an engine is developed using samples from the same cultural, acoustical, and linguistic backgrounds, great accuracy may not be attained due to acoustic similarities amongst emotions. Due to factors such as the considerable influence of cultural background, recording environment, and language employed in emotional utterances, recognising emotion from speech is a tough process. As a result, developing a high- performance emotion identification engine is tough. Even when an engine is developed with samples from the same cultural, acoustical, and linguistic backgrounds, great accuracy may not be attained due to acoustic similarities between one emotion and another. To solve these challenges, hierarchical classifier engines have been developed that increase emotion identification accuracy despite differences in the speaker's cultural background, recording environment, language utilised, utterance text, and age and gender. This article has introduced a novel convolutional neural network and devised a feature selection algorithm based on the F1 measure to select a specific feature for each emotion class. To determine the resultant emotion, the weighted response of representative features is determined. The proposed CNN was trained and evaluated using five distinct corpora from various cultural backgrounds, including the EMO-DB, SAVEE, IITKGP-SEC, Spanish corpus S0329, and Woogles corpora. The proposed CNN architectures are relatively generic and can be modified to accommodate variations in the quantity and character of features.