PhD Seminar • Artificial Intelligence | Machine Learning • Mission Impossible: Post Editing to Reduce Errors in ASR Transcripts

Please note: This PhD seminar will take place in DC 2102.

Ankit Vadehra, PhD candidate
David R. Cheriton School of Computer Science

Supervisor: Professor Pascal Poupart

In this work we focus on post-editing Automatic Speech Recognition (ASR) generated transcripts for customer support calls in French and English. Considering the limited amount of training data and the high cost of manual transcription, we use a semi-supervised method to generate a large synthetic corpus for training sequence-to-sequence post-editing correction models. We also perform data analysis and annotation to categorize the different ASR error types, utilizing the error-tagged set for better model evaluation. Deriving task insights from the results of the baseline method, we suggest further improvements to the synthetic dataset creation process for ASR transcripts, resulting in an overall improvement in the quality of the French and English call transcripts through post-editing.

In this talk, I present insights into the various challenges to the task of ASR transcript error correction, obtained through data analysis, along with the results of the post-editing models trained using different semi-supervised approaches to reduce the errors in the ASR transcripts.