CMU Spoken Document Retrieval in Trec-8: Analysis of the role of Term Frequency TF.
Matthew Adam Siegler, Rong Jin, Alexander G. Hauptmann · 1999
Retrieval Track used the basic same Sphinx speech recognition system as in TREC-7. Due to some unfortunate defaults in the parameter setup files, the speech recognizer did not perform in a reasonable manner. We will not analyze the results of the speech recognizer runs, as we believe the results contained abnormal types of errors, and insights or improvements on these errors would not generalize. A thorough examination of the speech recognition condition is given in [3]. However, we did evaluate a slightly modified weighting scheme in the reference (R1) and baseline (B1) conditions, which is described below. 1. Motivation