Selecting an acoustic correlate for automated measurement of /r/ production in children

Abstract

A current need in the field of speech pathology is the development of reliable and efficient techniques for the purpose of evaluating changes in speech production over the course of treatment. The industry standard for scoring speech is time consuming and expensive, as it involves aggregating perceptual ratings across expert listeners. As techniques for automated measurement of speech improve, acoustic measures have the potential to play an expanded role in the clinical management of speech disorders. The current study asks which of several acoustic measures of children’s productions of English /r/ corresponds most closely with ratings given by trained listeners. This study fits a series of ordinal mixed effects regression models to a large sample of children’s /r/ productions that had previously been rated by three trained listeners (speech-language pathologists). Controlling for age, sex, and allophonic contextual differences, the acoustic measure that accounted for the most variance in speech rating was F3-F2 distance, normalized relative to a sample of age- and gender-matched speakers. Therefore, this acoustic measure is recommended for use in future automated scoring of children’s production of rhotic targets.