Formant frequency estimation of speech signals plays an important role in speech synthesis, compression, and recognition. For example, formant information serves as a significant acoustic feature and offers a phonetic reduction in speech recognition. It plays a vital role in the design of some hearing aids [1]. Free resonances of the vocal-tract (VT) system are called formants. Formants are associated with peaks in the smoothed power spectrum of speech. Among different formant estimation techniques, linear predictive coding (LPC) based methods have received considerable attention [2]. In this case, formant frequencies are computed from the autoregressive (AR) parameters of the VT system. Most of the formant frequency estimation methods, so far reported, deal only with noise-free environments. However, formant estimation from noisy speech signals is difficult but an essential task as far as practical applications are concerned. In order to handle noisy environments, recently in [1], a method based on an adaptive band-pass filter-bank, later referred as AFB method, has been proposed where the estimation accuracy depends on initial estimates.
展开▼