Post-Editing Automatic Speech Recognition Error Correction

This research tackles the problem of correcting errors produced by automatic speech recognition systems when transcribing calls by customers to call centers. These transcripts are increasingly analyzed using automated natural language processing tools, however the quality of this analysis is highly dependent on the quality of the transcription. An automatic speech recognition (ASR) system is typically used to transcribe conversations into text, but despite the use of commercial ASR, the word error rate remains high for English and for French agent-customer calls at Intact Financial. Hence, there is a need for techniques that can reduce errors in a post-processing step.
This research will consist in the development of algorithms for post ASR error correction in French and English. This research will reduce the word error rate in the transcription of Intact Financial’s call transcription, which will in turn enable a better analysis of the calls for improved call quality, agent performance and customer satisfaction.

Faculty Supervisor:

Pascal Poupart

Student:

Partner:

Intact

Discipline:

Computer science

Sector:

Finance and Insurance

University:

University of Waterloo

Program: