Lattice Re-Scoring During Manual Editing for Automatic Error Correction of ASR Transcripts

Automatic speech recognition (ASR) systems are increasingly used to transcribe text for publication or official uses. However, even the best ASR systems make mistakes that can change the meaning of the recognition results. The results from these systems are therefore often reviewed by human editors, who fix the errors that arise. Offering automatic updates of utterances, with lattice re-scoring, could decrease the manual labor needed to fix errors from these systems. The research presented in this paper is conducted within an ASR-based transcription system with human post-editing for the Icelandic parliament, Althingi, and aims to automatically correct down-stream errors once the first error of a sentence has been manually corrected. After manually correcting the first error of the utterances, a new path is computed through the correction, using the lattice created during the ASR decoding process. With re-scoring, the sentence error rate (SER) for utterances containing two errors (and hence with SER=100%) drops to 82.77% and for utterances containing three errors drops to 95.88%. This paper demonstrates that the trade-off between automatically fixed errors and new errors introduced in the re-scoring heavily favours adding this process to the transcription system.