Phonetic Speech Recognition Using Pitch-Based Spectra
Original Publication Date: 1988-Oct-01
Included in the Prior Art Database: 2005-Feb-15
This article describes a method for phonetic speech recognition using pitch-based spectra. Speech recognition usually relies on acoustical parameters which are computed on a fixed-duration frame basis. However, for voiced sounds, processing on a pitch period basis yields more accurate and stable parameters, namely normalized spectra. Since voiced sounds are more frequent than unvoiced ones, this allows for a better recognition rate. Given successive sampled speech signal values, named (sk), one computes a precise pitch period p, expressed as a non-integer multiple of the sampling time. An integer number P > p is chosen to create an interpolated signal frame. P can be selected in different ways, depending on implementation considerations for the Fourier Transform calculations.