Abstract:This paper presents a new framework based on eliminating falling trend for improving average magnitude difference function (AMDF) pitch detection algorithm in speech signal processing to conquer the defect of the conventional AMDF which may generate error pitch. In the framework, linear polynomial is employed to represent the falling trend of AMDF. Then a modified AMDF algorithm based on least square (LS) method within the proposed framework is showed. LS method is performed to fitting the linear polynomial based on AMDF and the modified AMDF called LSAMDF is obtained by subtracting the falling trend, namely, the linear polynomial from AMDF for pitch detection. The simulated results show that the performance of LSAMDF is much better than that of many improvements of AMDF and validate the correctness of the proposed framework and the rationality of linear polynomial based falling trend representation.