Performance of acoustic phoneme models varies in a phoneme-based recognition system. Some phonemes may be modeled very well, providing generally higher acoustic likelihood scores; phonemes that are not modeled well will typically produce poorer scores. To calibrate the diverse performance of the phoneme models, an acoustic confidence score (ACS) has been derived which uses probability distributions of acoustic likelihood scores associated with each of the phoneme models in a recognition system to determine an objective measure of acoustic match. Preliminary results correlate the ACS of a hypothesized sentence and its correctness. One potential application of ACS is in the task of rejection of a recognized sentence due to its poor score. A sentence containing many of the phonemes that were not modeled well in the system will unfairly get a poor overall score when using the raw acoustic scores, and may be falsely rejected. ACS provides an objective measure of acoustic match, independent of the recognized phonemes.
Bibliographic reference. Rivlin, Ze'ev (1995): "A confidence measure for acoustic likelihood scores", In EUROSPEECH-1995, 523-526.