Spoken Document Retrieval (SDR) Track
Task: ad hoc search of news broadcasts
- documents:
- 550 hours of audio from TDT 2 corpus
- approx. 21,500 news stories, but commercials and fillers retained this year
- different transcript types
- reference transcript: close captioning, manual, ROVER reconciliation of automatic recognitions
- a baseline transcript produced at NIST using BBN�s Rough �N Ready recognizer (last year�s B2 transcript)
- participant�s own recognizer
- transcripts produced by other participants