An Ear to the Web

“The decline of america is dual to the automatic transmission home which actually came way back and go for the peace made a sleazy made us lazy.” Modern poetry? Actually, it’s a transcript of National Public Radio’s popular “Car Talk” program generated by SpeechBot, an experimental engine for interpreting and searching audio content on the Web created at Compaq’s Cambridge Research Laboratory (www.compaq.com/speechbot).

SpeechBot downloads audio files from the Internet, then uses speech-to-text software to make a transcript, says Compaq researcher Pedro Moreno. Although the software is specially tuned to decipher typical Internet file types like RealAudio, variable quality and a large number of different speakers keep accuracy low-the Car Talk guys’ thick Boston accents and tendency to talk over one another may make recognition tougher than usual. Even in this example, though, enough keywords came through to make the search worthwhile.