Robust feature extractors for continuous speech recognition

This paper presents robust feature extractors for a continuous speech recognition task in matched and mismatched environments. The mismatched conditions may occur due to additive noise, different channel, and acoustic reverberation. In the conventional Mel-frequency cepstral coefficient (MFCC) feature extraction framework, a subband spectrum enhancement technique is incorporated to improve its robustness. We denote this front-end as robust MFCCs (RMFCC). Based on the gammatone and compressive gammachirp filter-banks, robust gammatone filterbank cepstral coefficients (RGFCC) and robust compressive gammachirp filterbank cepstral coefficients (RCGCC) are also presented for comparison. We also employ low-variance spectrum estimators such as multitaper, regularized minimum- variance distortionless response (RMVDR), instead of a discrete Fourier transform-based direct spectrum estimator for improving robustness against mismatched environments. Speech recognition performances of the robust feature extractors are evaluated in clean as well as multi-style training conditions of the AURORA-4 continuous speech recognition task. Experimental results depict that the RMFCC and low-variance spectrum-estimators-based robust feature extractors outperformed the MFCC, PNCC (power normalized cepstral coefficients), and ETSI-AFE features both in clean and multi-condition training conditions.