Speech Emotion Recognition Using Cepstral Features Extracted With Gammatone Filter Banks Realized Based on ERB and Mel Frequency Scales

Speech emotion recognition (SER) involves identifying a speaker’s emotional state from their speech utterance. Prior research has explored various cepstral features for developing SER systems. Among these, Mel-frequency cepstral coefficients (MFCC) and gammatone cepstral coefficients (GTCC) are widely used. MFCCs and GTCCs are extracted using Mel filter banks and gammatone filter banks (GTFB), respectively. Traditionally, GTFBs are designed based on the equivalent rectangular bandwidth scale. Alternative frequency scales can be employed to design different variants of GTFBs, yielding distinct GTCC feature variants. In this paper, a novel Mel-scale-based GTFB (GTFB-M) is introduced to extract a novel variant of GTCC features, termed GTCC-M. The mathematical framework for designing GTFB-M is presented. The proposed GTFB-M together with the conventional Mel and gammatone filter banks are used to extract three types of cepstral features from emotional speech signals in the Berlin Database of Emotional Speech (Emo-DB) and the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS). To evaluate SER performance, speaker-dependent (SD) and speaker-independent (SI) SER systems are developed using support vector machines with one-versus-one and one-versus-all strategies. The effectiveness of individual and combined cepstral features is analyzed. Experimental results show that the GTCC-M features derived from GTFB-M perform comparably to traditional MFCCs and GTCCs in emotion recognition. Moreover, combining the proposed features with the conventional cepstral features enhances the overall SER performance. In the case of the Emo-DB database, standalone GTCC-M features achieves recognition (testing) accuracies of 80.12% (SD) and 63.03% (SI). The best recognition (testing) accuracies of 88.82% (SD) and 79.17% (SI) are achieved using an optimal feature combination, i.e., MFCC + GTCC + GTCC-M, when compared to the independent use of MFCC and GTCC features. Similarly, for the RAVDESS database, standalone GTCC-M features achieves recognition (testing) accuracies of 61.84% (SD) and 40.83% (SI). The best recognition (testing) accuracies of 78.62% (SD) and 46.88% (SI) are achieved using an optimal feature combination, i.e., MFCC + GTCC + GTCC-M, when compared to the independent use of MFCC and GTCC features. These findings underscore the potential of GTCC-M features to improve the SER performance when integrated with the MFCC and GTCC features.