Speech recognition as an input tool for electronic medical record enables efficient data entry at the point of care. We evaluated the speech recognition accuracy of IBM ViaVoiceTM for doctor-patient dialogues and for pronounced medical vocabularies. The recognition accuracy for doctor-patient dialogues was 95.4%, while that for pronounced medical vocabularies was 55.1%. In order to put speech-based electronic medical record to practical use, mis-recognized vocabulary must be significantly corrected. This paper describes a correction system for mis-recognized medical vocabulary for speech recognition-enabled electronic medical record. The correction system is composed of an extraction and a correction steps. In the extraction step, hamming distance between a parsed substring and the nearest medical vocabulary in the vocabulary database greater than 50% of the length of the substring was used to determine if the substring is a possible mis-recognized medical vocabulary. In the correction step, possible mis-recognized medical vocabularies are scored such that when both the code and location of a syllable is the same with those of a medical vocabulary found in our database, +5 is given and when the code is the same but the location is not, +1 is given. The medical vocabulary with the highest score in the database is used as the correction for the mis-recognized one. When 33 patient-doctor dialogues with 33 medical vocabularies were tested for three times by six testees (i.e., 33 × 6 × 3 = 594 sentences), 94% of the mis-recognized words were correctly detected and repaired. Poor recognition performance for hard medical vocabularies can be markedly improved by the mis-recognized medical vocabulary correction system. |