The talking cure

IF YOU have sausage-sized fingers, find pen-driven handheld computers a fiddle or have never got the hang of predictive text on your mobile phone, a new chip might provide a sympathetic ear. It is being devised by a team of researchers from Carnegie Mellon University and the University of California at Berkeley to do one thing, and one thing only: speech recognition. Using a new, hardware-based approach to the problem, the researchers hope to create a chip that performs speech recognition much more efficiently than is currently possible using software-based recognition systems. If they are right, it might soon become possible to dictate an e-mail into your BlackBerry, or edit your mobile phone's address book using voice commands alone.

Speech-recognition software has been on the market for over a decade, and in the past five years it has become advanced enough to displace keyboard entry, for some users at least. But speech-recognition packages such as IBM's ViaVoice and ScanSoft's Dragon NaturallySpeaking require a powerful desktop computer. Ask a portable device to do the same kind of computational heavy-lifting, however, and its battery will be flat within minutes. Why would a chip-only solution be any better?

The reason is simple: doing something in software is more flexible, but doing the same thing with a dedicated chip consumes far less power. Computationally difficult tasks often start out in software, and are implemented in hardware later. “You do them in software first, because it's easier,” says Rob Rutenbar, professor of electrical and computer engineering at Carnegie Mellon and the lead engineer on the “In Silico Vox” speech-chip project. “You redo them in hardware later to maximise their performance.”