Comparative performance of pitch detection algorithms on dysphonic voices

Current pitch detection algorithms run into difficulties when used on dysphonic voices. Two major sources of difficulty are the presence in the phonatory output of frictional, non-harmonic energy (in whispery voices), and microperturbatory fundamental frequency jitter and amplitude shimmer (in harsh and creaky voices). For adequate performance on dysphonic voices, pitch detection algorithms should have the following characteristics: 1. work on acoustic recordings from men, women and children 2. be noise resistant 3. work on continuous speech. Measures of pitch perturbation are defined. Three pitch detection algorithms were applied to the speech of dysphonic speakers as well as a control group of speakers. Two detectors work in the time domain (simplified inverse filter tracking (1) and a parallel processing method (2)), and one in the frequency domain (cepstral pitch detection (3)), Their comparative performance on perceptually rated clinical material is discussed.