Abstract: The performance of most current speech recognition systems degrades markedly innoisy environment. So, how to improve the robustness of system is the focus ofthe current speech recognition. The wavelet to the pre-process of the noisy speech recognition is applied, distinguishes the voiced from unvoiced speech in thewavelet domain, and deals with the voiced and unvoiced speech with threshold differently, and then performs the recognition after the inverse-transformation of denoising wavelet coefficients. The result demonstrates that the wavelet process is promising to the performance of noisy speech recognition.

Key words: speech recognition, wavelet, thresholding method