Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
A new approach for the analysis and evaluation of the pathological voices is proposed. The analysis tools include the following algorithms: robust pitch period (To) detection, high resolution pitch-synchronous spectral analysis, deconvolution using the cepstrum (c(t)), separation of harmonics from other spectral components, calculation of the smoothed spectrum corresponding to the vocal tract by means of the negative of the first derivative of the phase of the signal termed the group delay function (GDF) and calculation of the bispectra (B(fl,f2)) of the harmonics. Pitch synchronous spectral analysis is realized to evaluate c(t), GDF and B(fl,f2) with minimal error. Low quefrency windowed c(t) (corresponding to the vocal tract impulse response) is used for the GDF calculation. The harmonics are evaluated using the value calculated for To and spectral thresholds are adjusted for the low and high spectral regions found by means of GDF. B(fl,f2) is calculated only over the harmonics in order to reduce computation and facilitate the interpretation of the results. The effects are being quantified and the new measures should help the evaluation and classification of pathological voices.
Bibliographic reference. Boyanov, B. / Chollet, Gérard (1992): "Pathological voice analysis using cepstra, bispectra and group delay functions", In ICSLP-1992, 1039-1042.