Correcting mis-recognitions using basic travel expression corpus

摘要

We aim at improving the performance of our speech recognition system by detecting and correcting mis-recognitions. In this paper, we propose a new error correction method based on large basic travel expression corpus. This method has three steps: 1) Searching similar examples, 2) Extracting hypothesis words from examples, 3) Rescoring hypothesis words by using phonetic distance. Our experiments show that more than 20% of mis-recognitionsare corrected. In this paper, we also report some analyses of error correction results.