The Role of Speech Input in Wearable Computing

89 S peech recognition seems like an attractive input mechanism for wear-able computers, and as we saw in this magazine's first issue, several companies are promoting products that use limited speech interfaces for specific tasks. However , we must overcome several challenges to using speech recognition in more general contexts, and interface designers must be wary of applying the technology to situations where speech is inappropriate. On a tram in Zürich the other day, my research group started to discuss a counter-terrorism scenario I'm writing for the US Defense Advanced Research Projects Agency (DARPA). Going up a hill, the tram became quite noisy, so naturally I adjusted my voice to compensate. The tram stopped, and I suddenly found myself yelling, in very clear English, " Unless, of course, you are an American and get kidnapped by al-Qaeda. " All conversation on the tram ceased. Although amusing, this story is no doubt familiar. Almost everyone has had a similar experience in a crowded restaurant or at a cocktail party. We speak differently in the presence of noise: increased amplitude, reduced word rate, and clearer articulation. 1 The effect is sometimes called Lombard speech, after the French physician who noted it in 1911. Unfortunately, Lombard speech represents a difficulty for using speech recognition on a wearable computer. Most commercial speech recognizers are trained for dictation in an office environment and fail miserably when used in a noisy mobile environment. Although noise-canceling microphones help considerably with reducing noise level, the speaker's change in voice due to ambient noise means that even speech systems trained specifically to understand the user could have problems. In a recent talk, Chalapathy Neti from IBM showed results that begin to address the first problem of mobile speech recognition , that of background noise. 2 IBM tested its speech recognition system in conditions where it kept adding louder " speech babble " background noise to the clean speech to be recognized. The clean speech had a 19.5-decibel signal-to-noise ratio (SNR), and the recognition engine performed at a 12 percent word error rate on this speech. When researchers added additional noise to achieve an SNR of 12 dB, the word error increased to 60 percent. By 0 dB, word error approached 100 percent! Training the system on speech that was degraded by the same level and source of noise helped considerably: under 30 percent error at 12 dB and 80 …