A new spectra] envelope estimation method which uses FO-adaptive smoothing and compensation of logarithmic power spectrum is proposed for improving approximation accuracy around spectral peaks. In this paper, simulations using MRI-based vocal tract area functions and a glottal flaw model were conducted for tuning internal parameters, and evaluating estimation accuracy. Itakura-Saito distance with auditory-based frequency weighting was used as the evaluation index of spectral distance between the simulated target envelope and the estimates. Comparative test results indicated that the proposed method yields the best (smallest) spectral distance among conventional methods such as spectrogram, cepstrum, LPC spectrum and legacy-STRAlGHT.
展开▼
