Principal-components analysis for low-redundancy encoding of speech spectra

Abstract

The principal-components statistical procedure for data reduction is used to efficiently encode speech power spectra by exploiting the correlations of power spectral amplitudes at various frequencies. Although this data-reduction procedure has been used in several previous studies, little attempt was made to optimize the methods for spectral selection and coding through the use of intelligibility testing. In the present study, principal-components basis vectors were computed from the continuous speech of several male and female speakers using various nonlinear spectral amplitude scales. Speech was synthesized using a combination linear predictive (LP) principal-components vocoder. Of the amplitude scales investigated for use with a principal-components analysis of speech spectra, logarithmic amplitude coding of non-normalized spectra emerged as a slight favorite. Speech synthesized from four principal components was found to be about 80% intelligible using a form of the Diagnostic Rhyme Test for rhyming word pairs and about 95% intelligible for words within a sentence context. Speech synthesized from spectral principal components compared favorably in intelligibility and quality with speech synthesized from a control LP vocoder with the same number of parameters.