In this paper, the use of the probabilities produced by a Knearest neighbours (K-nn) estimator as confidence measure is investigated in an hypothesis verification post-processing scheme. The objective is to classify as correct or incorrect the outputs of a Gaussian mixture model (GMM) / HMM speech recognition system. Four confidence measures based on the K-nn probability estimator are introudced. Preliminary experiments are reported and discussed on the TIMIT database.
展开▼