An analysis of natural vs. synthetic speech intelligibility

Abstract

With Veterans Administration support, Haskins Laboratories has been developing a method of speech synthesis production for automatic reading aloud of printed text, with the goal of applying this technique to a practical reading machine for blind people. The laboratory prototype, as it exists today, uses for input a low cost Optical Character Recognition (OCR) device capable of reading (i.e., recognizing the print of) typewritten pages. The machine-readable orthographic text created by the OCR reader is then processed by a dictionary program which converts the input words to phonetic form. This program assigns stress and intonation symbols according to rules based on word type, context and sentence punctuation. The resulting phonetic code is then made visible to an editor who can insert corrections, if deemed necessary, before synthesis of the sentences begins. (Eventually the program will operate with no editorial intervention.) A series of intelligibility tests have been administered to both blind and sighted students at the University of Connecticut* in circumstances which allowed comparison of their listening performances with synthetic speech and—with natural speech. The tests, which are still in progress, have yielded results which indicate that the perception of synthetic speech places somewhat heavier demands on a listener's language processing capacity than does natural speech. However, this increased load appears to interact strongly with the subject content of the material, the syntactic structure, the punctuation provided in the text, and the speaking rate used in the output. An analysis of the results of this continuing evaluation study will be presented at the Conference.