Speech technology based assessment of dysarthric speech: preliminary results
Gwen Van Nuffelen, Catherine Middag, Jean‐Pierre Martens, Marc S. De Bodt · Ghent University Academic Bibliography (Ghent University) · 2007
Purpose: One of the objectives of the SPACE-project (Speech Algorithms for Clinical and Educational Applications) is to develop a speech technology based clinical assessment that provides reliable quantitative analyses of pathological speech.Method: Four automatic speech processing systems were applied on monosyllabic word recordings of a standardized Dutch phoneme intelligibility assessment.Systems 1 and 2 are automatic word recognizers which provide word accuracy rates.System 1 (WAR-ACF) was supplied with standard acoustic features, system 2 (WAR-ARF) with articulatory features, derived from the acoustic features.Systems 3 (CS-ACF) and 4 (CS-ARF) are automatic speech aligners that determine the best alignment between a speech sample and its canonical phonetic transcription.Confidence scores (CS) are computed for each phoneme (system 3) or for each articulatory feature (system 4).These CS are finally converted into a global score, designed to maximally agree with the perceptual intelligibility score.Samples of 60 dysarthric speakers were analyzed objectively by the four systems and perceptually by an experienced speech-language-pathologist.Pearson correlation coefficients (r) between the objective and the perceptual intelligibility scores are estimated by means of 5-fold cross validation experiments.Results: The correlations for systems 1 and 2 were respectively found to be moderate (r:0.56) and low (r:0.33).However, the alignment-based systems resulted in much higher correlations (system 3: r:0.72; system 4: r:0.72).Conclusions: Alignment-based systems, provide more reliable intelligibility scores than recognition-based systems.No significant difference was found between working with acoustic and working with articulatory features.The current results are encouraging but further refinements are needed.