(A) comparative performance study of pitch extraction algorithms for noisy speech

Abstract

Pitch detector that makes voiced/unvoiced (v/uv) decision of input speech and determines pitch period of voiced speech is an essential and important device in speech processing systems. In this thesis a performance comparison study of several pitch detection algorithms for noisy as well as clean speech has been done. To do the comparative study, a data base of six sentences spoken by a male and a female has been prepared, and their pitch periods and v/uv decisions were determined for reference data by the eye detection method. The algorithms detect pitch periods and v/uv decisions were compared with those reference data. The performance criteria of each algorithm are the number of voiced decision errors, unvoiced decision errors and pitch errors. In addition, based on the results of the performance comparison, a new algorithm that has been improved from the autocorrelation method with center clipping has been developed. This new pitch detector yields better performance than others so far proposed.