Objective Measure of Speech Intelligibility Using Linear Predictive Coding.

摘要

Four distance measures of speech intelligibility based on linear predictive coding (LPC) are developed and evaluated. The data base used for evaluating the measures consisted of lists of 58 words from Diagnostic Rhyme Test IV. The lists were transmitted over a spread spectrum radio communications channel and subjected to 7 different levels of non-white, non-Gaussian jamming noise. The lists were all scored subjectively for intelligibility by a trained listener panel. The subjective scores were used to judge the effectiveness of the four distance measures. The Articulation Index was also calculated for each of the word lists and compared to the LPC measures as to effectiveness and efficiency in measuring speech intelligibility. The Articulation Index was significantly more effective than the LPC measures. The best LPC measure provided 42% correlation with the subjective scores. The Articulation Index provided 69% correlation. The overhead associated with data tape alignment and parameter computation makes LPC measures extremely inefficient as compared to the Articulation Index. (Author)