The Development of a Method for Evaluating and Predicting Speaker Recognizability in Voice Communication Systems

Although standardized tests for evaluating the effects of voice coding and communication systems on speech intelligibility and speech "quality" are presently available, a standard method of evaluating the effects of such systems on speaker recognizability is yet to be developed. A major obstacle to this end is the difficulty of controlling listeners' familiarity with voices used in such tests, in particular the difficulty of obtaining and maintaining listening crews whose members have equal familiarity with a speaker sample that is large enough to provide adequate representation of the population of interest. Dynastat proposes to develop a method in which listeners' familiarly with voices is not a factor. It is based on the simple premise that recognition presupposes discrimination-- discrimination with respect to the perceived voice traits (PVT's) in terms of which voices differ among themselves. The proposed method would predict speaker recognizability in communication systems by evaluating the effects of those systems on listeners' ability to discriminate or resolve differences among voices with respect to various PVT's-- i.e., the effects of the system on the speaker-identify information carried by the various PVT's. NEcessary steps involve determining the nature and number of PVT's that carry significant amounts of speaker-identify information and the development of rating scales by means of which listeners can reliably evaluate voices with respect to these PVT's. The final step involves the experimental test of various practical and economical methods of evaluating system-induced loss of speaker-identity information via the PVT's.