Reconstructing Medical Dictations from Automatically Recognized and Non-Literal Transcripts with Phonetic Similarity Matching

In this paper, we describe the automatic reconstruction of literal transcriptions for medical dictations from a non-literal transcription and an automatically recognized speech transcript by phonetic similarity matching and alignment. We present a customized phonetic similarity measure which is trained on a set of phonetically similar string pairs, returns interpretable alignment results, and is robust in its application. Furthermore, we introduce flexible automatic phonetic transcription with regular expressions to deal with formatted entities in written texts and alternative pronunciations in recognized texts. In an evaluation, our method reduced the word error rate for the reconstructed transcription by 12% relative.