Noise-Robust Pitch Detection Based On Biorthogonal Wavelet Transform and Formant-Frequency Tracing Algorithm

Abstract

In this paper, an improved wavelet-based pitch determination method is presented. Based on the biorthogonal wavelet transforms and a new formant-frequency tracing algorithm, the proposed method can accurately and efficiently extract the pitch information from the speech signal even for noisy environments. In this new method , the input speech will be first decomposed into a formant-contain approximation signal using the fiorthogonal wavelet transforms and the formant-frequency tracing algorithm. Then, the pitch information can be detected in the formant-contain approximation signal via an object-depended maximum extractor and a pitch correction algorithm. Besides the robustness against noise, the proposed method does not embed any segmentation window and predefined threshold which are usually needed in the conventional wavelet-based pitch detection algorithms. From the experimental results, one can find that the proposed method works well for synthetic and natural speech signals. In addition, it is also shown that the proposed signals. In addition, it is also shown that the proposed algorithm can achieve the better performance than those of the present wavelet-based as well as the conventional time-domain and spectral-domain pitch detectors under noisy conditions.