Meta-models for confidence estimation in speech recognition

We describe an approach to confidence estimation that attempts to decouple the contributions of the acoustic and language model components to speech recognition output. The output of the acoustic models when decoding phonemes is itself modelled using HMMs to produce a set of models which we term meta-models. When benchmarked against a “standard” method for assigning confidence (the N-best score), the meta-models gave a relative improvement of 6.2%. Furthermore, it appears that the N-best and meta-models techniques are complementary, because they tend to fail on different words