Error Heuristic Based Text-Only Error Correction Method for Automatic Speech Recognition

With the fast development of deep learning, automatic speech recognition (ASR) has achieved significant improvement. However, there still exists some errors in ASR transcriptions. They will greatly interfere with the downstream tasks, which take the transcribed text as source data. Obviously, it is necessary to set a corrector to reduce errors in the ASR transcriptions. For various downstream tasks, a text-only based corrector would be more adorable because of its minimal requirements for recognizer. However, the limitation of ASR decoding information exerts considerable influence on the performance of corrector. Correcting a correctly recognized word into a wrong one is one of the most common problems. To relieve this situation, we propose to adopt error knowledge from an error detection model as a heuristic to train a sequence-to-sequence (Seq2seq) correction model by transfer learning. In this way, the corrector can focus on correcting the wrong words with a soft label. The experiment shows our method can effectively correct ASR errors, with a 4.35% word error rate (WER) reduction for the transcription. It outperforms the state of the art Seq2seq baseline with a 1.27% WER reduction.