Alternative hypothesis generation using a weighted kernel feature matrix for ASR substitution error correction

Abstract

Although automatic speech recognition (ASR) has been successfully used in several applications, it is still non-robust and imprecise especially in a harsh environment wherein the input speech is of low quality. Robust error correction for ASR outputs thus becomes important in addition to improving recognition performance. In recent approaches to error correction, linguistic or domain information is used to generate the alternative hypotheses for the ASR outputs followed by the selection of the most likely alternative. In this study, the distances between ASR outputs and the potentially correct alternatives are estimated based on a weighted context-dependent syllable cluster-based kernel feature matrix followed by multidimensional scaling (MDS)-based distance rescaling. These distances are then used to construct an alternative syllable lattice and the dynamic programming is used to obtain the most likely correct output with respect to the original ASR results. Experiments show that the proposed method achieved about 1.95% improvement on the word error rate compared to the correction pair approach using the MATBN Mandarin Chinese broadcast news corpus.