Hunnicutt, 1984). Each symbol stands for a common word. The authors found that Bliss symbols seem to be a good way to get nonvocal children started on language production, but a switch to normal orthography seems desirable later.

Vocal communication difficulties are fairly common when many prelingual deaf try to communicate with a hearing individual who is unfamiliar with the speech of the deaf. Bernstein et al. (1984) have designed a telephone communication system to overcome this problem. A deaf person can speak utilizing a Speech Plus text-to-speech board, and can "listen" by viewing the output of a large-vocabulary isolated-word speech recognition device. Preliminary data suggest that the performance of the recognition system in current use is marginal, but still good enough to be useful; future improvements could make such systems more attractive. An important issue is how to format the recognition alternatives, e.g., phonemes or a word hypothesis lattice (Huggins et al., 1986).

B. Training aids

In one sense, a talking aid is by default a language training aid because it promotes practice and elicits direct feedback. This is one reason why it would be advantageous to get more talking aids to nonvocal children as early as possible in their school career. Experience suggests that this kind of device will also promote correct spelling and syntax (Carlson et al., 1980). The inherent attraction of computer devices may mean that the approach could also be used with normal children for initial reading instruction.

A novel and quite successful application of text-to-speech is in the area of training dyslexic children to read. Dyslexia is a self-perpetuating difficulty because it is embarrassing to be helped by a teacher or friend, and it is nearly impossible to practice reading without help. Now several research groups have devised computer systems that permit unsupervised reading practice (Atkinson, 1972; Olson et al., 1985). For example, the system being developed at the University of Colorado-Boulder (Olson et al., 1985), uses a computer display screen, a mouse pointer, and a DECtalk text-to-speech system to read unfamiliar words or sound them out syllable by syllable.

Training aids need not be restricted to handicapped individuals (Sherwood, 1981). It is well known that speech has measurable advantages over reading and writing in many cognitive situations (Ochsman and Chapanis, 1974). For example, Suppes (1979, 1981) devised a computerized course for teaching algebra, and showed that providing some of the interactions via spoken responses resulted in better learning performance than visual presentation of all computer responses. Nakatani et al. (1986) provide spoken tutoring in the use of a text editor, noting that otherwise the student must constantly switch attention between the behavior of the editor and any tutorial information provided at the bottom of the screen. Tutorials of this sort can be made to have quite natural intonation and phrasing by proper annotation of the tutorial text (Hirshberg and Pierrehumbert, 1986).

C. Reading aids for the blind

Another application area is the development of reading aids for the blind. According to a survey by the National Center for Health Statistics (NCHS, 1977), approximately 1.4 million Americans are so severely visually impaired that they are unable to read ordinary news print, even with glasses. Machines that can scan printed material and produce speech would be of great help to this community. Ultimately, the goal is personal reading machines, although the current cost of the best performing machine, Kurzweil's, at a price of over $30 000, is far from this objective.

With the advent of computer typesetting, and the large text data bases available to computer users, it may not be necessary for a blind person to obtain a high-cost text reader. In some cases, connecting a text-to-speech system to a personal computer that is interfaced to a large network may serve many information gathering needs. One interesting pilot project in Sweden uses an FM overnight broadcast to load the day's newspaper into a blind person's personal computer, and has indexing programs to permit scanning of topics, using the Infovox SA-101 text-to-speech system (Carlson et al., 1976; Carlson et al., 1981).

Other efforts in this area have been concerned with timely production of talking books for the blind. Currently, most cassette recordings of books for the blind are produced by Recording for the Blind in Princeton, New Jersey. They use volunteer readers, and find that it takes up to 6 months after an order is placed to produce an audio copy of a typical textbook -- primarily because of the problems inherent in coordinating volunteer efforts when so many hours of speech are to be produced. There are two possible ways to speed up the delivery of textbook orders. A text-to-speech system can work day and night to produce audio cassette tapes, or the text could be placed on a disk for a personal computer -- in which case a blind person having a personal text-to-speech system could listen to the book and potentially be able to skim and scan over the book much more efficiently than is presently possible. An adjustable speaking rate, in conjunction with a computerized index or other method of content addressing could make reading almost as easy as the browsing we take for granted when we pick up a book (example 36 of the Appendix).

Another application is in the area of aids in the workplace. Of the 1.4 million visually impaired, many are elderly. However, 37 000 are children below 18, and 360 000 are between 18 and 64. Of these, about 106 000 are employed, according to the National Center for Health Statistics. Those blind individuals who work in an office environment could increase their productivity and become less dependent on a sighted co-worker if some sort of "talking text editor" computer system were available. Currently available systems are reviewed in Aids and Appliances Review (McGillivray, 1983).

D. Medical applications

Most medical applications are no different from other business applications, in that large health maintenance organizations employ centralized computer-based records on