Invited Talk: Processing Broadcast Audio for Information Access

Processing Broadcast Audio for Information Access Jean-Luc Gauvain, Lori Lamel, Gilles Adda, Martine Adda-Decker, Claude Barras, Langzhou Chen, and Yannick de Kercadio Spoken Language Processing Group LIMSI-CNRS, B.P. 133, 91403 Orsay cedex, France (gauvain@limsi.fr http://www.limsi.fr/tlp) Abstract This paper addresses recent progress in speaker-independent, large vocabulary, continuous speech recognition, which has opened up a wide range of near and mid-term applications. One rapidly ex- panding application area is the process- ing of broadcast audio for information access. At LIMSI, broadcast news tran- scription systems have been developed for English, French, German, Mandarin and Portuguese, and systems for other languages are under development. Au- dio indexation must take into account the specificities of audio data, such as needing to deal with the continuous data stream and an imperfect word tran- scription. Some near-term applications areas are audio data mining, selective dissemination of information and me- dia monitoring. 1 Introduction A major advance in speech processing technology is the ability of todays systems to deal with non- homogeneous data as is exemplified by broadcast data. With the rapid expansion of different me- dia sources, there is a pressing need for automatic processing of such audio streams. Broadcast au- dio is challenging as it contains segments of vari- ous acoustic and linguistic natures, which require appropriate modeling. A special section in the Communications of the ACM devoted to “News on Demand” (Maybury, 2000) includes contribu- tions from many of the sites carrying out active research in this area. Via speech recognition, spoken document re- trieval (SDR) can support random access to rel- evant portions of audio documents, reducing the time needed to identify recordings in large multi- media databases. The TREC (Text REtrieval Con- ference) SDR evaluation showed that only small differences in information retrieval performance are observed for automatic and manual transcrip- tions (Garofolo et al., 2000). Large vocabulary continuous speech recogni- tion (LVCSR) is a key technology that can be used to enable content-based information access in au- dio and video documents. Since most of the lin- guistic information is encoded in the audio chan- nel of video data, which once transcribed can be accessed using text-based tools. This research has been carried out in a multilingual environment in the context of several recent and ongoing Euro- pean projects. We highlight recent progress in LVCSR and describe some of our work in de- veloping a system for processing broadcast au- dio for information access. The system has two main components, the speech transcription com- ponent and the information retrieval component. Versions of the LIMSI broadcast news transcrip- tion system have been developed in American En- glish, French, German, Mandarin and Portuguese. 2 Progress in LVCSR Substantial advances in speech recognition tech- nology have been achieved during the last decade. Only a few years ago speech recognition was pri- marily associated with small vocabulary isolated word recognition and with speaker-dependent (of- ten also domain-specific) dictation systems. The same core technology serves as the basis for a range of applications such as voice-interactive database …

Content truncated to 500 words