What is “confidence measures”?
evaluate reliability of recognition results.
Why measure confidence or reliability?
ambient noises, speaker variations, channel distortions and other mismatches.
Major methods in category?
- combination of predictor features;
- word duration
- word graph density
- posterior probability: estimate the probability of the sequence given the acoustic observations.
- hypothesis testing: likelihood ratio = (word likelihood score) / (background model + competing model)
- utterance verification.
non-keyword rejection or keyword recognition: detect if the input speech does not contain any of the words in the recognizer vocabulary set.
- Out-Of-Vocabulary Detection and Confidence Measures for Speech Recognition Using Phone Models.
- Confidence Estimation for Machine Translation