| Popis: |
SRI has been evaluating automatic speech recognition for text input by speech‐impaired people. The speech of 85 speech‐impaired people (primarily deaf and cerebral palsy) was recorded. This speech was input to a high‐performance, template‐based speech recognizer. With a 300‐word vocabulary, the machine recognized approximately 25% of the impaired speakers better than human listeners did (in context); with a 50‐word vocabulary, the machine recognized approximately 57% better than listeners. A preprototype text‐entry and communication aid is being evaluated in our laboratory. Twelve subjects have tried this system briefly, four of whom could use speech input faster than typing input after only a few minutes of practice. Two subjects (who were not immediately faster with speech input) were trained and evaluated on the system more extensively. Both subjects' performance with the speech recognizer improved dramatically; after 9 h of practice, one's speaking rate almost doubled his manual typing rate. This presentation will describe this system, discuss our results, and show videotapes of the two trained users. [Work supported by NIH.] |
