SPEECH TECHNOLOGY BASED ASSESSMENT OF DYSARTHRIC SPEECH G
SPEECH TECHNOLOGY BASED ASSESSMENT OF DYSARTHRIC SPEECH G. Van Nuffelen, SLP M. De Bodt, Ph. D C. Middag, Ir J. P. Martens, Ph. D SPACE-PROJECT SPEECH ALGORITHMS FOR CLINICAL AND EDUCATIONAL APPLICATIONS Antwerp University Hopsital University of Antwerp University of Ghent IALP August 2007 Copenhagen
Purpose Develop a speech technology based clinical assessment that provides reliable analyses of pathological speech. Quantitative: objective intelligibility score Qualitative: segmental, articulatory errors Preliminary results for dysarthric speech IALP August 2007 Copenhagen
Methods: speech samples • audio records • 60 dysarthric speakers • DIA: phoneme level • 50 test items (CVC) – subtest A (19): initial consonants – subtest B (15): final consonants – subtest C (16): medial vowels & diphthongs • 25 equal versions IALP August 2007 Copenhagen
Methods: speech samples top IALP August 2007 Copenhagen
Methods: speech samples • perceptual score: percentage correctly identified phonemes • judgments: experienced SLP • strong inter-rater (ICC: 0. 91) and intra-rater reliability (ICC: 0. 93) IALP August 2007 Copenhagen
Methods: speech processing systems 4 systems: 1) WAR-ACF 2) WAR- ARF 3) CS- ACF 4) CS- ARF ACF: acoustic features ARF: articulatory features WAR: speech recognizer CS: speech aligner IALP August 2007 Copenhagen
Methods: ACF • systems 1 & 3 • ± 1000 statistical acoustic models (HMMs) • triphone models • acoustics of a phoneme when appearing in conjunction with a particular preceding and succeeding phoneme • co-articulation IALP August 2007 Copenhagen
Methods: ARF • systems 2 & 4 • 25 binary ARF derived from acoustic features • voicing, vowel height, manner of articulation, place of articulation, … • 40 models – 40 phonemes IALP August 2007 Copenhagen
Methods: speech recognizer • systems 1 & 2: automatic word recognizers • each item DIA: lexicon: fixed frame + all possible target phonemes (~ listeners strategies) • log likelihood score/ word lexicon ‘perceived’ • Word Accuracy Rate: percentage of correctly identified words • linear regression model: WAR ‘objective’ score IALP August 2007 Copenhagen
Methods: speech aligner • each item corresponding word & canonical (= expected) phonetic transcription known • segmentation in time intervals – Acoustic realizations (system 3) – Phoneme components (system 4) IALP August 2007 Copenhagen
Methods: speech aligner • confidence scores – / phoneme (system 3) – / articulatory feature (system 4) • probability that the phoneme/ articulatory feature did occur where expected? • linear regression model: CS ‘objective score’ IALP August 2007 Copenhagen
Methods: comparison perceptual and objective intelligibility scores • Pearson correlation coefficients • 5 -fold cross validation experiments • 5 separate data sets • 4/5: training data • 1/5: test set • mean r IALP August 2007 Copenhagen
Results IALP 1) WAR-ACF: r: 0. 56 2) WAR-ARF: r: 0. 33 3) CS- ACF: r: 0. 72 4) CS-ARF: r: 0. 72 August 2007 Copenhagen
Results IALP August 2007 Copenhagen
Discussion • alignment-based: more reliable • reliability of perceptual assessments – DIA: inter-rater: r=0. 91 – AIDS (Yorkston, 1984): inter-rater: r=0. 92 • CS-ACF – CS-ARF: no difference for quantitative analyses • ARF: segmental, articulatory errors IALP August 2007 Copenhagen
Thank you for your attention! gwen. van. nuffelen@telenet. be This research was granted by the Flemish Institute for the Promotion of Innovation by Science and Technology (IWT). IALP August 2007 Copenhagen
WP 5: future • Segmental, articulatory analysis • Release of the assessment tool feedback from clinicians • Release of the speech corpus – Put data and metadata together – Validation – TST-central
- Slides: 17