| Summary: | 碩士 === 國立臺灣科技大學 === 資訊管理系 === 95 === Average Magnitude Difference Function (AMDF) is a pitch detection algorithm that is known to often suffer from the problem of pitch misjudgments. Various algorithms have thus been proposed to address this problem. However, as most of these algorithms focus on how to improve the procedure of AMDF itself, we argue that extracting the Local Minima pitch points may also play a critical role on enhancing the pitch detection accuracy, after applying the original AMDF algorithm.
This paper proposed a pitch point extraction algorithm which improves over the traditional AMDF algorithm, and it consists of three steps. First, we apply the AMDF algorithm to calculate the value of each transformed point, locate all Local Minima points, and discard those points whose values are greater than a threshold. Second, we compare the remaining Local Minima points with their neighboring points and retain only the smallest values around the Local Minima points. Finally, we determine the Local Minima fundamental frequency, under the constraint that the fundamental frequency discrepancy rate must be lower than a pre-specified threshold determined from the chromatic scale discrepancy rate.
Compared with the original version of AMDF, which may erroneously determine the pitch from the smallest Local Minima point, this proposed method could offer more accurate results. In spite of the additional overheads due to the processing of Local Minima points, the resulting accuracy is improved, while the overall calculation time is still well acceptable.
|
|---|
