Investigation of different time–frequency representations for detection of fricatives

Fricatives are an important class of speech sounds which exhibit noisy characteristics with dominant high-frequency components. Accurate detection of fricative segments in continuous speech is useful in developing hearing aids as persons with hearing impairment have difficulty in perceiving high-frequency sounds. This work investigates time–frequency representations for accurate detection of fricatives in continuous speech. Time–frequency representations with fixed and progressive temporal resolution are investigated and temporal resolution is analyzed. Spectrograms are computed using S-transform, constant-Q transform and single frequency filter approach. Single frequency filtered spectrogram exhibits the best temporal resolution among the three. Feature highlighting the frication characteristics is derived from these time–frequency representations and a threshold based approach is used to detect the fricatives. Experiments use utterances from American English TIMIT database. Fricative detection is a prior step to modify the pathological speech to improve the intelligibility of pathological speech. The system is also evaluated for pathological speech from Torgo database. Experimental results show that feature derived from constant-Q transform and single frequency filter can detect fricatives accurately compared to the feature derived from S transform.