Inter-speaker comparison of LPC acoustic space using a minimax distortion measure

A method is described that measures the range of frequency spectra produced when a person speaks. Our "minimax" method selects a certain size subset from a set of LPC vectors such that the maximum RMS log spectral distance from any vector in the superset to its nearest neighbor in the subset is minimized. The minimum inter-vector distance in the subset is, in general, monotonically related to the volume of the region that contains the superset. Passages of English, Japanese, and Spanish were recorded, and analyzed using this method. Results indicate that the range of spectra characteristic of different individual\´s speech are significantly different. There is less difference, if any, between languages in their measured acoustic ranges. The results are discussed with reference to issues in phonology and speech coding.