Computer Speech Recognition as an Objective Measure of Intelligibility

摘要

The effectiveness of computer speech recognition as an objective measure of intelligibility was explored. A commercially available program was assessed following the manufacturer's protocol to analyze speech samples from three speakers without neuro-genic speech disorders, one speaker with moderate spastic dysarthria, and a synthesized speech sample. The system identified the synthesized speech most accurately, and the speech of the speaker with dysarthria least accurately. To improve clinical relevance, a variation of the recommended protocol was performed. The first author provided a referent speech sample, and typically used intelligibility tests produced by three speakers with dysarthria were assessed. These results were compared with intelligibility judgments from a large number of everyday listeners. Compared to the everyday listener results, the software judged all samples much less accurately. Single word intelligibility was particularly disparate because of the prediction model used by the recognition system. In addition, replicability was poorer than expected, again, due to the prediction model. Further work, modifying the software's referent sample as well as assessing other recognition programs, is proposed. These programs have the potential for providing reliable pre-post therapy analysis and a stable point of comparison across clinicians and settings.