A front-end for emotional speech classification based on new sub-band filters

Emotional state of speaker is among many information included in the speech signal that is ignored in Automatic speech recognition (ASR). An automatic emotion detection system bears the task of extracting appropriate features that could represent the emotional characteristics of the speaker. This study is focused on finding more effective features in frequency domain for speaker emotional state classification. Here, we replace Mel filters with new filter banks, designed according to the data and the importance of emotional components. This approach is based on Kullback-Leibler divergence (KLD) and F-ratio. The new filter banks have led to an improvement in the emotion recognition system performance of about 3.2% on average, in comparison with the baseline system, mel frequency cepstral coefficients (MFCC).