Project : metiss
Keywords : speech, audio, signal, analysis, processing.
A broadcast news indexing system
Participant : Guillaume Gravier.
The concept of rich transcription consists in generating an orthographic transcription of a show enriched with side information concerning the speakers, the presence of background music, the topics and any information related to the structure of the show.
METISS, in collaboration with the Computer Science Dept. at ENST Paris, developed a radio broadcast news rich transcription system that was evaluated during the dry-run phase of the national evaluation campaign ESTER (Broadcast News Rich Transcription System Evaluation). This system will also be evaluated in the official test campaign of ESTER, scheduled early 2005.
In the 2004 dry-run evaluation, METISS participated in the following tasks : transcription, audio event tracking and speaker tracking. The transcription was based on the Sirocco speech recognition software. Audio event tracking permitted the validation of results previously obtained in the framework of the Domus Videum project, concerning segmentation and simultaneous event detection. Finally, the speaker tracking system was adapted from the system used in the last NIST evaluation campaigns on speaker recognition.
In 2004, a huge effort has been put on the development of this system which is now fully operational. The system is used to validate the Sirocco speech recognition engine and our know-how in speaker characterisation. Some new approaches and results produced by the research group (as reported in the next section) were incorporated to the IRENE system.
Contact : firstname.lastname@example.org