Parameters for Automatic Speaker Recognition over LPC (Linear Predictive-Coding) Systems

摘要

This investigation focused on the selection of acoustic parameters related to speaker identity for possible use in automatic speaker recognition systems designed to operate with speech processed through low data rate linear predictive-coding (LPC) voice communication systems. Parameters selected were: formant frequencies and bandwidths (based on both spectral and line-spectrum frequency analyses), average and standard deviation of pitch, spectral tilt, and averages of LPC reflection coefficients. Preliminary recognition results indicate that individually these parameters have potential; combinations of parameters should prove even more useful, particularly if less robust aspects of the measurements are eliminated from the recognition decision.