Statistical Error Correction Methods for Domain-Specific ASR Systems

摘要

Whenever an ASR company promises to deliver error-proof transcripts to the end user, manual verification and correction of the raw ASR transcripts cannot be avoided. This manual post-editing process systematically generates new and correct domain-specific data which can be used to incrementally improve the original ASR system. This paper proposes a statistic, SMTbased ASR error correction method, which takes advantage of the past corrected ASR errors to automatically post-process its future transcripts. We show that the proposed method can bring more than 10% WER improvements using only 2000 user-corrected sentences.