VT-AMDF, a pitch detection algorithm

Abstract

This paper proposes a VT-AMDF pitch detection algorithm which based on average magnitude difference function (AMDF). Time processing is important when implement in real time system. In order to decrease computational time and complexity of pitch detection, the VT-AMDF uses only 55% of total time intervals. The method can reduce the computational time 1.3, 1.4, 4.8 and 5.4 times, compared with original AMDF, autocorrelation function (ACF), normalized square difference function (NSDF) and YIN, respectively. The experiments evaluate on seven words in five different tones of Thai isolated word. The results show that the gross error of VT-AMDF is 3.88%, which is more than original AMDF 0.48 and YIN 0.22, but less than ACF 0.54 and NSDF 1.47.