Studying the Effect of Frame-Level Concatenation of GFCC and TS-MFCC Features on Zero-Shot Children's ASR

The work presented in this paper aims at enhancing the recognition performance of zero-shot children's speech recognition task through frame-level concatenation of two complementary front-end acoustic features. The acoustic features chosen are TANDEM-STRAIGHT-based Mel-frequency cepstral coefficients (TS-MFCC) and Gamma-tone frequency cepstral coefficients (GFCC). The GFCC model the cochlear response of the human auditory system. The MFCC features, on the other hand, model the human pitch perception. Therefore, the GFCC and TS-MFCC features capture the acoustic information differently and that too with very low correlation. Consequently, concatenation of TS-MFCC and GFCC feature vectors helps in modeling complementary and a wider range of relevant acoustic information. This, in turn, enhances the recognition performance significantly. The experimental evaluations presented in this paper show that a relative reduction of nearly 12% is achieved by feature concatenation.