Abstract:
The very important point of applications of speech recognition in a real field is its robustness against noise disturbances. A new robust algorithm is proposed, which has the following three features: (1) It uses only one microphone for input and noise-free speech as references; (2) it simulates a human hearing process in the following three functions: frequency analysis by the critical bandwidth, physical to sensory level transformation, and lateral inhibition; and (3) the speech is enhanced by passing a noisy one to a filter made of the spectrum envelope derived by the linear predictive analysis of the noisy speech itself. Outputs of the analyzing filter are vector-quantized and recognized by VQ and HMM composed of the noise-free reference speech processed in the same way as the noisy ones. The method can improve recognition scores from 15% to nearly 30% in the range of SNR of 15--10 dB for additive white random noise, compared to those of the noisy speech by HMM of noise-free speech.