Periodicity Detection Based on Instantaneous Frequency

Periodicity is commonly found in many physical phenomena. This paper describes the notion of instantaneous frequency amplitude spectrum (IFAS) which originally developed in the context of time-frequency signal analysis. The IFAS provides better the harmonic structure representation of speech signal than the short time Fourier transform (STFT) amplitude spectrum. We define harmonicity measure as a quantity that indicates the strength of periodical regularity of the signal and that shows substantial difference between periodic signal and noiselike waveform. Speech signal is natural signal in which periodic (voiced) and random (unvoiced) component are coexisted. We show several voicing decisions based on harmonicity measure. We also describe a new technique for voicing decision based on harmonicity measure with variable window length and IF band selection. The threshold technique is adapted according to the harmonicity measure in unvoiced region which related to noise level of the input to determine a new preset of the threshold. The performance of the proposed method is evaluated with Japanese continuous speech signals in both clean and adverse environment. The results showed that the IFAS-based method outperformed autocorrelation-based, amplitude magnitude difference (temporal-based) and TEMPO (instantaneous frequencybased) methods.