Using clinician annotations to improve automatic speech recognition of stuttered speech

In treating people who stutter, clinicians often have their clients read a story in order to determine their stuttering frequency. As the client is speaking, the clinician annotates each disfluency. For further analysis of the client's speech, it is useful to have a word transcription of what was said. However, as these are realtime annotations, they are not always correct, and they usually lag where the actual disfluency occurred. We have built a tool that rescores a word lattice taking into account the clinician's annotations. In the paper, we describe how we incorporate the clinician's annotations, and the improvement over a baseline version. This approach of leveraging clinician annotations can be used for other clinical tasks where a word transcription is useful for further or richer analysis.