Speech recognition research at AT&T Bell Laboratories

Abstract

At AT&T Bell Laboratories, a broad range of systems for speech recognition, depending on the intended application area, has been studied. These systems have been extensively studied for isolated word (and phrase) recognition for command and control applications where a single word (or phrase) suffices to effect some type of control over a system or service. Typical applications include call type recognition for automated dialing and call routing within an organization. Several systems have also been implemented for connected word recognition with vocabularies ranging from the digits up to several hundred words. Typical applications for this technology include order entry, credit card entry, and digit dialing. Most recently, several studies have been begun of large vocabulary (1000–50 000 words), continuous speech recognition. The intended application for this type of system is database management and access. Our primary focus has been in studying alternative representations of subword speech units. In particular, recognition systems based on phones (acoustic‐phonetic units), diphones, and acoustically defined units have been implemented and studied. Several methods of representing and accessing words from subword units including lexical access from a stored dictionary, explicit network representations of words, and derived acoustic lexicons have been studied. An investigation using a parser and a covering grammar to represent the constraints of the language and tack on the spoken input has also been shown. In this talk, the progress of each of these areas of speech recognition research is illustrated. Also discussed are the long‐term research goals and the steps intended to take to reach these goals.