When the Recognizer Listens

It's not only what the recognizer hears that can pose a challenge, it's also when it listens. Computers ”like people ”need to know when to start listening and when to stop listening.

Have you ever had someone start talking to you when you were deep in thought? You probably missed some of what they initially said. That's because you weren't prepared to start listening. Likewise, sometimes when you cut off someone who is speaking before he or she finishes a thought, you might miss some important piece of information. Sensing when to start and stop talking in a conversation is easy for most of us. But it's more difficult for computers because lots of things can sound like speech to them. To a computer, any sound ”a cough, music playing, a slamming door, a crying baby ”could be someone trying to communicate with it. It's only after analyzing the sound that the computer can determine whether or not it was speech.

In general, speech recognizers also need to know when a person has finished talking. It would be unfortunate if a caller started to say, "I'm leaving from Boston and going to San Francisco," but the computer cut her off after the first half of the sentence , then asked, "OK, and where are you going?" Conversely, it would also be frustrating for there to be an exceptionally long pause before the computer said something in response.

For all of these reasons, it's important for designers to understand the limitations and strengths of their speech recognizer ”and to design the system accordingly .