Spectral-temporal receptive fields and MFCC balanced feature extraction for noisy speech recognition

摘要

This paper aims to propose a new set of acoustic features based on spectral-temporal receptive fields (STRFs). The STRF is an analysis method for studying physiological model of the mammalian auditory system in spectral-temporal domain. It has two different parts: one is the rate (in Hz) which represents the temporal response and the other is the scale (in cycle/octave) which represents the spectral response. With the obtained STRF, we propose an effective acoustic feature. First, the energy of each scale is calculated from the STRF. The logarithmic operation is then imposed on the scale energies. Finally, the discrete Cosine transform is applied to generate the proposed STRF feature. In our experiments, we combine the proposed STRF feature with conventional Mel frequency cepstral coefficients (MFCCs) to verify its effectiveness. In a noise-free environment, the proposed feature can increase the recognition rate by 17.48%. Moreover, the increase in the recognition rate ranges from 5% to 12% in noisy environments.