A confidence measure for speech recognition systems based on two maximum entropy approaches

We implement a confidence estimation stage for a speech recognition system by using two maximum entropy models. These models takes information from various sources including a set of scores which have proved to be useful in confidence estimation tasks, and efficiently combine those scores. Two different approaches of the maximum entropy model are developed. First a basic model which takes advantages of smoothing techniques used in a previous work, and second an optimized model, which is designed to hold a set of very few but essential characteristics of the model. Resulting model not only increase the performance of the system, but also reduces dramatically the number of parameters. We take advantages of maximum entropy paradigm by imposing restrictions to the model which are otherwise very difficult to model by using classical maximum likelihood estimation. Both models are evaluated with two different corpus and compared to a model previously developed. Keywords—Confidence estimation, maximum entropy, confidence measures, speech recognition.