Year: 2008
Type: conference-paper
Abstract: Intelligibility is a primary measure for the assessment of pathological speech. Traditionally, it is measured using a perceptual test, which is by definition subjective in nature. Consequently, there is a great interest in reliable, automatic and therefore objective methods. This paper presents such a method that incorporates an automatic speech recognizer (ASR) for producing features that characterize the pronunciations of a speaker and an intelligibility prediction model (IPM) for converting these features into an intelligibility score. High correlations (about 0.90) between objective and perceptual scores are obtained with a system comprising two different speech recognizers: one with traditional acoustic models relating acoustical observations to triphone states and one using phonological features as an intermediate layer between the acoustical observations and the phonetic states.
Source: Interspeech (USB)
Institutions: Ghent University, University of Antwerp
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W84239132
