Robust voice command interface design through the use of error templates

A scheme for enhancing the robustness of conventional M-ary pattern classifiers in noisy, cluttered, and distorting environments by the addition of error templates is presented. Independent error templates implement a nonparametric clutter model and are effective in reducing false positive errors in the presence of structured disturbances, whereas linked error templates reduce category errors (falsely labeling data from category i as category j). To test the effectiveness of independent and linked error templates and to determine any problems that they may create, digital simulations were done. A hardware prototype using error templates was built and used by a handicapped client with speech and motor disabilities. The simulations show that the addition of error templates significantly decreases false positives, with some improvement in the other categories. Results for the hardware prototype are encouraging.