A hybrid pitch period (To) detector characterized by parallel analyses of the speech signal in temporal, spectral and cepstral domains and preprocessing for periodic/aperiodic(unvoiced) separation (PAS) is proposed. The preprocessing is realized by analyses in these three domains and PAS by multilayer perceptron neural network. To evaluation is carried out by analyses in: 1. Time domain 1.1 Preprocessing by adaptive central clipping (ACC). 1.2. Calculation of the short-term autocorrelation function (Rst(T)) 1.3. Calculation of To by analysis of the distances between the maximas in Rst(t) - LADM. 2. Spectral domain: 2.1. Spectral ACC. 2.2. Calculation of Rst(x) over the power and logarithmic spectra. 2.3. To detection by LADM. 3. Cepstral domain: 3.1. Littering the cepstra to eliminate the formant structure. 3.2 Calculation of Rst(x). 3.3. To detection by LADM. Final To evaluation by a logical analysis of the results from the three domains. EXPERIMENTS Two phonations of the vowel "a" of 40 speakers and 62 patients were analyzed. For the proposed detector errors were significantly minimized.
Cite as: Boyanov, B., Hadjitodorov, S., Teston, B., Doskov, D. (1997) Robust hybrid pitch detector for pathological voice analysis. Proc. LARYNX 1997, 55-58
@inproceedings{boyanov97_larynx,
author={B. Boyanov and S. Hadjitodorov and Bernard Teston and D. Doskov},
title={{Robust hybrid pitch detector for pathological voice analysis}},
year=1997,
booktitle={Proc. LARYNX 1997},
pages={55--58}
}