If transcripts were already available, then these were aligned to the audio stream; otherwise, an approximate transcript was produced using speech recognition. The Calista recognizer that was used was derived from Sphinx-3. Due to the low quality of streaming audio at the time, the word error rate was quite high, but most searches were still able to retrieve relevant hits.[7] The search results linked to the offset in the stream that corresponded to the search phrase, so that users did not need to listen to the entire program to find the section of interest.
Eberman, B.; Fidler, B.; Iannucci, R.A.; Joerg, C.; Kontothanassis, L.; Kovalcin, D.E.; Moreno, P.; Swain, M.J.; Van Thong, J-M (March 1999). "Indexing Multimedia for the Internet". Compaq Technical Report. CRL 99/2. Archived from the original on March 20, 2006.
Dufaux, F.; Eberman, B.; Kontothanassis, L.; Moreno, P.; Swain, M.; Weikart, C. (March 1999). "A system for indexing web multimedia". Compaq Technical Report. CRL 99/3.