Comparison of subjective and objective speaker recognition under voice disguise conditions

An experiment was performed in order to compare the results of subjective and objective
speaker recognition under voice disguise conditions. The experimental material consisted of
the key sentence “To jest akustyka” (Eng. “This is acoustics”) spoken several times by Polish
male speakers in a natural mode and under voice disguise conditions. In the subjective method
the utterances were grouped in pairs and presented to the listeners whose task was to make
a decision whether a given pair of speech samples was produced by one speaker or by two
different speakers. In the objective method two parametrical representations of speech (LPC
coefficients and Km reflection coefficients) were utilized and a computer program for automatic
speaker verification was applied. For normal speech both methods exhibited a very good
effectiveness of speaker recognition and the results for the subjective method were a bit lower
(98.9% in comparison to 99.4% for the objective method). Under voice disguise conditions,
however, the mean results for the subjective method (92.3%) were substantially better than for
the objective one (77.2%).