Identification of speech transients using variable frame rate analysis and wavelet packets.

Abstract

Speech transients are important cues for identifying and discriminating speech sounds. Yoo et al. and Tantibundhit et al. were successful in identifying speech transients and, emphasizing them, improving the intelligibility of speech in noise. However, their methods are computationally intensive and unsuitable for real-time applications. This paper presents a method to identify and emphasize speech transients that combines subband decomposition by the wavelet packet transform with variable frame rate (VFR) analysis and unvoiced consonant detection. The VFR analysis is applied to each wavelet packet to define a transitivity function that describes the extent to which the wavelet coefficients of that packet are changing. Unvoiced consonant detection is used to identify unvoiced consonant intervals and the transitivity function is amplified during these intervals. The wavelet coefficients are multiplied by the transitivity function for that packet, amplifying the coefficients localized at times when they are changing and attenuating coefficients at times when they are steady. Inverse transform of the modified wavelet packet coefficients produces a signal corresponding to speech transients similar to the transients identified by Yoo et al. and Tantibundhit et al. A preliminary implementation of the algorithm runs more efficiently.