Robust Speech Feature Extraction Using the Hilbert Transform Spectrum Estimation Method

Abstract

The performance of traditional mel-frequency cepstral coefficients (MFCC) speech feature extraction method decreases drastically in the complex noisy environment. To improve the performance and robustness of speech recognition system, which is based on spectral envelope estimation method, the minimum distortionless response spectrum MVDR-MFCC (Minimum Variance Distortionless Response-MFCC) feature extraction method was proposed. However, the computational complexity of MVDR-MFCC is very high. In this paper, we proposed MHCC (Hilbert-MFCC) feature extraction method for speech, which introduced the Hilbert transform to MFCC process. The experiments, under 8 different noisy environments, indicate that, compared with MVDR-MFCC feature extraction method, the proposed method not only reduces the algorithm’s complexity significantly, but also is less affected by noises, achieving significant improvement in the robustness—the average recognition rate across different noise types and SNRs increases by 12%.