A Thesis Submitted to the Department of Information and Computer Science of Aalto University School of Science in Partial Fulfillment of the Requirements for the Degree of Licentiate of Science in Technology

A significant portion of the work time of dentists and nursing staff goes to writing reports and notes. This thesis studies how automatic speech recognition could ease the work load. The primary objective was to develop and evaluate an automatic speech recognition system for dental health care that records the status of patient's dentition, as dictated by a dentist. The system accepts a restricted set of spoken commands that identify a tooth or teeth and describe their condition. The status of the teeth is stored in a database. In addition to dentition status dictation, it was surveyed how well automatic speech recognition would be suited for dictating patient treatment reports. Instead of typing reports with a keyboard, a dentist could dictate them to speech recognition software that automatically transcribes them into text. The vocabulary and grammar in such a system is, in principle, unlimited. This makes it significantly harder to obtain an accurate transcription. The status commands and the report dictation language model are Finnish. Aalto University has developed an unlimited vocabulary speech recognizer that is particularly well suited for Finnish free speech recognition, but it has previously been used mainly for research purposes. In this project we experimented with adapting the recognizer to grammar-based dictation, and real end user environments. Nearly perfect recognition accuracy was obtained for dentition status dictation. Letter error rates for the report transcription task varied between 1.3 % and 17 % depending on the speaker, with no obvious explanation for so radical inter-speaker variability. Language model for report transcription was estimated using a collection of dental reports. Including a corpus of literary Finnish did not improve the results. Date: 30.3.2012 Language: English Number of pages: 70