5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

A New Confidence Measure Based on Rank-Ordering Subphone Scores

Qiguang Lin, Subrata Das, David Lubensky, Michael Picheny

IBM T. J. Watson Research Center, USA

This paper presents a new approach to measuring how confidently a word has been correctly recognized, or confidence measure. The approach consists of three major steps: (1) standard decoding; (2) forced Viterbi alignment (needed for stack decoders); and (3) rank-ordering of the subphone scores. More specifically, from the aligned sentence the third step computes the likelihood scores of the hypothesized subphone and all other competing subphones. A list of the subphones is generated in the descending order of the scores and a rank is assigned to the hypothesized subphone according to its positioning. Additional processing of selective weighting and upper-bound limiting is applied to minimize contamination of rank computations by bad segments or by highly-variable phones. The obtained rank is then used as the confidence measure. Results of word rejection experiments show that the new approach outperforms other measures such as whole-word scores by reducing the equal error rate from 32% to 20%.

Full Paper

Bibliographic reference.  Lin, Qiguang / Das, Subrata / Lubensky, David / Picheny, Michael (1998): "A new confidence measure based on rank-ordering subphone scores", In ICSLP-1998, paper 0806.