Phoneme adjustment in enhanced speech

A system has been developed to enhance the quality and intelligibility of noisy, mutilated speech. The system processes speech in the frequency domain using a 512-point DFT (discrete Fourier transform). The amplitude spectrum of voiced regions of speech is smoothed in order to reduce the effects of noise. Harmonics of the glottal pitch frequency of voiced speech and peaks of amplitude spectrum for unvoiced speech are selected with a nonlinear rule to further reduce and eliminate the nonspeech components from the spectrum. The harmonics selected are not necessarily the exact harmonics of the glottal frequency. Speech is then reconstructed using the amplitude, phase, and frequency of the specific harmonics/peaks selected. The number of harmonics/peaks used for reconstruction varied from 40 to 50 in each frame of 256 DFT points. The reconstructed speech has much better subjective quality and improved signal/noise ratio