Human or Machine?—a Study of Voice Naturalness

Critical examination of human voiced speech indicates that both sustained vowels and connected speech often show considerable variability in the spacing of successive glottal pulses. Such fluctuations are too rapid to be reproduced by a low-frequency parametric pitch signal as classically used in vocoders. What is the perceptual significance of these fluctuations? A preliminary investigation of this question utilized sustained vowels spoken by two talkers. The fluctuations in these utterances were removed by digital computer processing. These samples were used, together with the originals, to form A-B pairs, each pair containing one “machine” and one “human” utterance. Subjects were presented with the A-B pairs and asked to identify the one made by the human. Naive subjects were able to perform almost perfectly in this task. Since rapid period-to-period fluctuations were the distinguishing characteristic of the human samples, this result is evidence for the role of such fluctuations in voice naturalness.