FREQUENCY-DOMAIN SPECTRAL ENVELOPE ESTIMATION FOR MONOPHONIC AND POLYPHONIC SIGNALS

摘要

Input speech is analyzed to extract frequency information. Peak-picking is used to identify formant frequencies of the psectral envelope according to psycho-acoustic laws. Differences are measured between successive spectral envelopes (502) and at each point, accumulations of the differences are compared to a threshold to determine whether the signal is varying slowly or quickly (504). If varying slowly, then a smoothing factor approaching one is assigned to u (506). If varying quickly, then a smoothing factor approaching zero is assigned to u (508). The signal is then smoothed according to S'(w) = uSn(w) + (1-u)S'n-1(w) (510). S'(w) designates a smoothed value, Sn(w) is the current signal, S'n-1 is the previous smoothed value, u is the weight factor and (w) indicates frequency domain.