Spoken Document Retrieval (SDR) Results
- Task: search 4 different versions of broadcast news transcripts
- 100 hours of broadcast news, approx. 3000 stories
- human transcript (reference)
- baseline speech recognizer transcripts, one at 35% SWER (baseline 1), one at 49% SWER (baseline 2)
- participant’s own recognizer (reco)
- transcripts produced by other participants (cross-recognizer)
- 23 topics created for track
- standard trec_eval evaluation
- compare robustness of different retrieval technologies
- compare suitability of different recognizer technologies