Jpetiot/ janvier 8, 2007/ Previous

Exploring large sets of audio documents for extraction and processing of concersational speech

Problem description

Very large collections of speech data are now available and have to be indexed to allow later retrieval of recorded information. The cost of manual transcription of audio recordings is high, especially when specific indexing is wanted such as the main topic, keywords or the name of the speakers. Automatic rich transcription can be done at a reasonable cost, but the error rate of the systems has to be as low as possible to allow efficient exploitation of their outputs.


The goal of project EPAC is to propose methods for information extraction and document structuring and signature which would be specified to audio data and applied to a huge set of audio documents (1930 hours of recorded data). These methods will take into account all the information channels present in those data and will analyze them from different granularity levels : signal segmentation (speech/music/jingle/etc.), environment characterization (crowd, interview, debate…), speaker identification and tracking, speech transcription, topic detection and tracking, speech analysis, conversational interactions, speaker opinion, etc.


The EPAC project will focus more particularly on conversational speech processing, from detection to content analysis (who is speaking, about what…), this research field representing important stakes for the transcription and the study of spoken language. Because, among the various TV or radiophonic shows, the part of conversational speech is often minimal and overlooked. Methods for detecting and extracting conversational speech will be proposed and developed. The project will set up processing techniques for this particular kind of speech through the proposal of relevant descriptors and development of associated tools. Finally, a common evaluation framework for all project participants will be set up for each studied task. The evaluation will make use of the radio program database.


People involved in SAMOVA team



  • Start time : 1st january 2007
  • End time : 1st january 2010

Web page


Share this Post