Speech Features Enhancement Based on Frequency-domain ICA
LV Zhao1,2;WU Xiao-pei1; LI Mi2
1 The Key Laboratory of Intelligent Computing & Signal Processing Anhui University Hefei 230039)2 The First Aeronautical College of Air-Force Xinyang 464000
Abstract: To suppress the interference of convolutive noise on speech features and improve the rate of speech recognition, a speech features enhancement algorithm based on frequency-domain ICA (Independent Component Analysis) is presented in this paper. In the proposed algorithm, noise short-time spectrum is estimated by the frequency-domain ICA algorithm, and then noise reduction is achieved by subtracting the estimated noise short-time spectrum from the noisy speech short-time spectrum to be enhanced in the Mel-scale filter bank domain. As a result, robust MFCC (Mel Frequency Cepstral Coefficients) are acquired. Simulation and real environment experiential results reveal that the recognition ratio of the proposed algorithm obtains the relative increasing of 38.2% and 35.8% compared with conventional MFCC, which reveal that the mismatch between training features and testing features in convolutive noise environment can be suppressed effectively.