Cepstral Derivatives in MFCCs for Emotion Recognition

Recent advancements and growing interest in developing affective interaction systems demand more efficient and robust solutions for emotion recognition. Most of the current emotion recognizer systems use Mel frequency cepstral coefficients (MFCCs) along with its temporal derivatives, and other static spectral features. In this paper, we propose a new dynamic feature called cepstral deltas/derivatives to improve emotion recognition performance. We have used two emotion speech datasets of different languages (German, and English) to study the performance of cepstral deltas. The MFCCs are used as a base feature to compare the performance of the proposed feature and traditionally used temporal deltas. The result shows that the addition of cepstral deltas with MFCCs improves the performance of specific emotions in different languages. The possible performance of cepstral deltas in other classifiers like support vector machine and deep neural networks is discussed.