Computer-generated speech

The need for communication between man and machine is increasing as the use of computers pervades the industrial and information processing worlds. Since speech is perhaps the most natural means of communication, computer speech output is well-suited to applications in which an operator's eyes and hands are busy monitoring situations. The Instrumentation and Controls Division of the Oak Ridge National Laboratory (ORNL) sponsored this project because of the potential need to use computer speech output in numerous applications throughout ORNL. The specific objectives were (1) to establish the development tools necessary to readily implement a speech output capability in a computer system and (2) to evaluate commercially available speech modules for quality of speech and flexibility.

This thesis reviews the essential aspects of speech synthesis and distinguishes between the two prevailing techniques: compressed digital speech and phonemic synthesis. It then presents the hardware details of the five speech modules evaluated. FORTRAN programs were written to facilitate message creation and retrieval with four of the modules driven by a PDP-ll minicomputer. The fifth module was driven directly by a computer terminal. The compressed digital speech modules (T.I. 990/306, T.S.I. Series 3D and N.S. Digitalker) each contain a limited vocabulary produced by the manufacturers while both the phonemic synthesizers made by Votrax permit an almost unlimited set of sounds and words. A text-to-phoneme rules program was adapted for the PDP-11 (running under the RSX-11M operating system) to drive the Votrax Speech Pac module. However, the Votrax Type 'N Talk unit has its own built-in translator.

Comparison of these modules revealed that the compressed digital speech modules were superior in pronouncing words on an individual basis but lacked the inflection capability that permitted the phonemic synthesizers to generate more coherent phrases. These findings were necessarily highly subjective and dependent on the specific words and phrases studied. In addition, the rapid introduction of new modules by manufacturers will necessitate new comparisons. However, the results of this research verified that all of the modules studied do possess reasonable quality of speech that is suitable for man-machine applications. Furthermore, the development tools are now in place to permit the addition of computer speech output in such applications.