Multiple-factor speaker identification of normal and distorted speech

Abstract

The purpose of this research was to examine the possibility that a set of speaker identification vectors were sufficiently robust to function adequately in the presence of speaker and system distortions. Twenty-five adult male talkers produced three repetitions (normal, stress and disguise) of an extended reading passage from which long-term speech spectra (LTS), speaking fundamental frequency (SFF) and speaking time (ST) were extracted for four consecutive 32 s intervals. The last three of the 32 s segments constituted a reference sample in a discriminant analysis procedure; the first segment was utilized as the test sample. Each vector was used alone and also in combination with other vectors. Distortions included limited passband as well as speech produced while the talker was under stress or attempting a disguise. While the described approach functioned adequately for the normal speaking conditions, no vector (singly or in combination) adequately differentiated talkers when speech was distorted.