Robust classification of speech based on the dyadic wavelet transform with application to CELP coding

This paper describes a new algorithm for the classification of telephone-bandwidth speech that is designed for efficient control of bit allocation in low bit-rate speech coders. The algorithm is based on the dyadic wavelet transform (D_yWT) and classifies each unit subframe into one of the three categories background noise/unvoiced, transients/voicing onsets, periodic/voiced. A set of three parameters is derived from the D_yWT coefficients, each giving a decision score that the associated class is active. Taking the history into account, a finite-state model controlled by these parameters computes the classifier´s decision. The proposed algorithm is robust to various types of background noise. In comparison with a classifier based on the long-term autocorrelation function, the D_yWT classifier proves to be superior. To evaluate its performance in CELP-type speech coders, a variety of excitation coding schemes with bit rates between 2200 and 4800 bit/s is investigated