Towards a phoneme-based predictive communication aid for nonspeaking individuals

摘要

The high incidence of literacy deficits amongst people with severe speechimpairments (SSI) has been well documented. Without literacy skills, people withSSI are unable to use orthographic-based communication systems effectively tocompose novel words and messages in spontaneous conversation. To address this problem, previous research has proposed the use of phoneme-based communication systems, which enable users to access a limited set of spoken phonemes (i.e. speech sounds). By combining sequences of phonemes, novel linguistic items can be constructed without knowledge of orthographic spelling.Despite this, phoneme-based communication systems have been an under-researched topic. The few published reports on existing phoneme-based systems have highlighted a number of usability issues, including poor communication rate, difficult access methods to target phonemes, and high learning demands.To address these issues, this research aimed to investigate whether prediction methods could be employed to facilitate phoneme entry and the word creation process. Although prediction methods have been extensively applied in orthographic-based communication systems for rate enhancement, very limited research has been conducted on the potential of such methods to improve the usability of phoneme-based communication systems.This thesis describes the development and evaluation of a novel phoneme-basedpredictive communication system. The system utilised robust statistical language modelling techniques to perform context-dependent phoneme prediction and phoneme-based word prediction. The usability of these phoneme-based predictivemethods was assessed through a series of evaluation studies, includingcomputational experiments with simulated interfaces, formative evaluations withsixteen non-disabled participants, and longitudinal case studies with two adults whohave cerebral palsy and limited literacy. Results of these evaluations demonstratedthat the predictive methods led to substantial improvements in user performance,both in terms of entry rate and accuracy. Data from a comparative evaluation with anonspeaking male adult showed that his entry rate increased from 1.14 words perminute (WPM) without prediction to 4.19 WPM with prediction, while his error ratereduced from 50.0% word error rate (WER) without prediction to 0.0% WER withprediction. Results of a comparative study with a nonspeaking female adult reported that her performance improved from 0.37 WPM (79.17% WER) without prediction to 2.89 WPM (0.0% WER) with prediction. In addition, positive results of the case studies evidenced the potential of phoneme-based predictive communication systems to effectively support nonspeaking individuals with literacy difficulties in generating novel linguistic items.