The effects of instructions and feedback on speaker consistency in a speech recognition framework
Linda A. Roberts, Jay G. Wilpon, Dennis E. Egan, Jean Bakk · The Journal of the Acoustical Society of America · 1985
Past research has demonstrated that a major problem in automatic speech recognition is speaker consistency. Instructions (suggestions of specific behaviors for consistent speech) and feedback (a visual display indicating the distance between each utterance and its template) were evaluated in a speaker-dependent recognition system as means of helping people speak more consistently. Utterances of 60 subjects were recorded using a 2 × 2 design. Dependent measures included the difficulty of forming templates, distances between utterances and templates, and recognition accuracy. For every dependent measure, instructions had a positive effect on speaker consistency but feedback had no reliable effect. When subjects were retested 2 weeks after initial training, their speech was less similar to the templates than on the day the templates were formed, but the instructional effect was still reliable. An extensive analysis of speech errors was carried out. Judges subjectively classified each utterance according to different error categories. These subjective ratings were then compared to recognition accuracies. The error analyses suggested which aspects of the instructions were effective and how future instructions might be improved.