Comparison of the intelligibility of predictor coefficient and formant coded speech

Abstract

This study describes the analysis and synthesis of speech using linear prediction and formant coding methods. A detailed comparison is made of speech synthesized using 12 predictor coefficients, six predictor coefficients, five formant frequencies and amplitudes, three formant frequencies and amplitudes, three formant frequencies and amplitudes with monotone pitch, and three formant frequencies with monotone pitch and amplitudes calculated by formula. The synthetic speech and the original speech are examined using Griffith's articulation test for rhyming minimal contrasts. Intelligibility scores are measured for each method of synthesizing speech, and confusion matrices are constructed for each method. Possible explanations for the observed confusions are investigated with the view in mind of suggesting improvements to speech analysis-synthesis systems.