Speech Analysis Lab: Research on Machine-Mediated Language
Matthew Bauer, assistant professor of linguistics and director of the Speech Analysis Lab, investigates speech intelligibility in public and commercial environments. More and more, humans encounter machine-mediated language—human or synthetic speech delivered through technology. Examples range from public transit announcements to drive-through windows, automated phone systems, and toys that talk. For people with low vision, encounters with prerecorded or synthetic speech also include bank teller machines, Web security services, and public transit kiosks, among many others.
The quality of machine-mediated speech varies widely. Unfortunately, guidelines from the Americans with Disabilities Act specify only that communications to any member of the public should be "effective." Surprisingly, there are few design guidelines or standards for the prerecorded or synthetic speech.
One goal of Bauer's Speech Analysis Lab is to determine the qualities of intelligible machine-mediated speech. The lab uses a variety of technology, including a sound booth, computer bank, ultrasound machine, and other recording devices. Machine-mediated speech from various sources is analyzed for features that are known to affect intelligibility, such as signal fidelity, speaking rate, intonation, and pitch. The idea is to develop more specific technical specifications for intelligible machine-mediated speech.
So far, the Speech Analysis Lab has produced two reports for an advocacy group for people with disabilities. It has also run a series of IPROs, including a spring 2010 study of audio CAPTCHAs (Completely Automated Public Turing Test to Tell Computers and Humans Apart). Visual CAPTCHAs present distorted text to a user, who is asked to type the text into an answer box. The idea is that humans, but not computers, are able to interpret the distorted text, thus preventing online forms from being accessed by spambots.
Audio CAPTCHAs ask users to identify an audio string of digits or words that has been distorted or placed against a background of "white" noise or reversed speech The audio format is intended to be accessible to users whose vision problems prevent them from using visual CAPTCHAS. At issue is how to improve audio CAPTCHAs so that users can easily solve them but computers cannot.
In the future, Bauer "hopes to secure a client base in order to analyze assistive speech before it goes out to the public, before it gets unfavorable reviews, and in certain extreme cases before lawsuits."posted: