Utilizing Open Vocabulary Word Vector for Detecting Errors in Decoded Speech Transcripts

Annotating newly collected audio data requires substantial resources and time. Employing existing pretrained automatic speech recognition (ASR) decoder to initially perform decoding on the data can significantly reduce the resources and time. Nevertheless, the decoding will introduce some errors to the transcripts. It is therefore beneficial to have an error detection model to detect which words are erroneous. Our proposed error detection model incorporates “intrinsic” property of the decoded words, in form of open-vocabulary word-vector, modifying an existing network architecture. The model is trained and optimized on the ASR decoding result of existing annotated audio data sets. Evaluation of the model using another distinct data set achieves a Matthews correlation coefficient (MCC) of 0.5846. The model can be employed further to accelerate the annotation process by predicting the decoding errors, which can be corrected either manually or automatically.