Navigating Audio Like Text: Browsing and Searching Voicemail Messages

(Joint work with Brian Amento, Michiel Bacchiani, Martin Jansche, Candy Kamm, Phil Isenhour, Meredith Ringel, Brian Roark, Aaron Rosenberg, Larry Stead, Steve Whittaker, Gary Zamchick)

Currently, large amounts of speech data are available in personal (voicemail), company proprietary (recorded conference calls, focus group sessions, customer care recordings), and public databases (newscasts, films). However, such speech corpora are difficult to make use of, for lack of tools to browse and search them. We have developed such tools in several audio browsing and search projects, one of which, SCANMail, I will describe in this talk. SCANMail combines automatic speech recognition with information retrieval and information extraction technologies to provide voicemail users with the ability to manipulate their voicemail messages much as they do their email messages, searching them by content, extracting key bits of information (phone numbers, callerids) from them, and accessing parts of them randomly rather than in sequence. Results are presented to the user in a client GUI whose capabilities have been tested in laboratory experiments and a field trial. Access via Java phone and in standard email clients is also possible. In the talk I will describe and demonstrate the system and discuss several novel features of the technology in more detail: the ranking of messages by urgency or private/business nature, using a combination of acoustic and lexical information, and the extract of key pieces of information from the automatic transcriptions. A demo of the SCANMail system is available at http://www.fancentral.org/~isenhour/scanmail/demo.html.