A new de-noising method based on Teager energy operator(TEO) and modified Wiener filter is proposed for improving the robustness of a speech recognizer in presence of additive noise. Voiced and unvoiced segment of speech exhibit contrasting characteristic. Therefore different methods are used for enhancing voiced and unvoiced speech. In unvoiced region the noise is reduced by modified Wiener filtering and feature is obtained such as mel-scaled frequency cepstral coefficients(MFCC). In voiced frame feature is obtained by removing the noise bias in the spectral domain, rectifying after the application of TEO and transforming the modified spectrum into cepstral coefficients in a fashion similar to MFCC.
We will report on experiments on Aurora2 DB and TI-digit DB contaminated by military noises. The experimental results show the proposed feature outperforms other methods for improving noise robustness.