Estimation of Fundamental Frequency of Noisy Speech Signals using Correlogram based on Subband Filtering

摘要

Estimation of the fundamental frequency (FO) is a difficult task under the colored noise environment. In this work, an algorithm for estimation of the fundamental frequency of speech signal is proposed based on multi-channel comb filtering and correlogram for noisy speech signals. The noisy speech signal is decomposed into four bands using wavelet transform. The comb filter is applied on each decomposed subband signal to enhance the peaks in each correlogram. For obtaining the correlograms, the autocorrelation is computed for each subband. The comb filter is defined using the raised cosine function with interpeak frequency ranges from 25 to 500 Hz. Comb filter channels are needed to match the harmonics of the fundamental frequency and these channels are selected based on the harmonic to subharmonic ratio. The proposed method is evaluated using the CSTR database. The performance of the proposed method is compared with the RAPT, YIN, Cepstrum and SHR algorithms in terms of gross pitch error (GPE). The simulation results show the performance of the proposed algorithm is superior in different noise environments compared to existing methods.