A two‐tap pitch predictor for measuring voice aperiodicity noise at high SNR

Abstract

The measurement of aperiodicity noise in the human voice is complicated by the multiple source of this noise. While turbulence noise at the glottis is hypothesized to contribute to aperiodicity noise, it is known that waveform jitter, the cycle‐to‐cycle fluctuation in pitch period, contributes significantly to the measured aperiodicity noise. In seeking to reduce the influence of jitter on measured noise, it has been observed that a cross‐correlation pitch tracker based on a sliding analysis window exhibits the greatest change in measured pitch period when the window crosses a glottal‐closure epoch. It was hypothesized that the accuracy of a pitch‐predictor based measurement of the noise could be improved if one knew the correct alignment of the analysis window relative to the glottal epoch. In the absence of this knowledge, for each update of a short‐term analysis, two alignments of the window were used and that alignment with the least‐square pitch predictor error was selected. This leads to a computationally efficient algorithm that does not require explicit determination of glottal epochs. Preliminary trials on subjects with high SNRs indicate a reduced influence of jitter on the observed SNRs.