Realtime recognition of unvoiced fricatives in continuous speech to aid the deaf

An algorithm to perform realtime recognition of unvoiced fricatives in continuous speech has been developed and tested. The tree-structured algorithm is implemented on a minicomputer, but it can also be economically realized by using simple digital hardware. Zero-crossing and amplitude information from successive 10 msec segments of the speech signal, its derivative and the outputs of two bandpass filters is used to isolate and identify the fricatives. Recognition of /s/, /θ/, /f/ and /
\int
/ is better than 95 percent correct with a total decision time of less than 90 msec. The algorithm has been successfully applied in a realtime system to aid the discrimination of unvoiced fricatives by persons with severe high-frequency hearing loss.