Using acoustic analyses of SLI children speech to study the effect of Language Interaction Therapy

Robin Peeters, Akos Steger, Gerda Bruinsma, Catia Cucchiarini and Helmer Strik

Approximately 7% of children worldwide have a specific language impairment (SLI). SLI interferes with the development of (first) language skills despite normal hearing, intellectual abilities and sufficient exposure to a language. Speech and language teachers can provide language therapy to children with SLI to improve their language skills. In The Netherlands, one of the approaches adopted for this purpose is called ‘Language Interaction Therapy’ (LIT).

LIT is described as “a focused stimulation intervention for children with weak morphosyntactic skills”. A study on LIT was conducted with 32 Dutch-speaking SLI children aged 4 to 6 years old. After a baseline phase of usual care (consisting of other therapies for morphosyntactic skills), participants either followed LIT or continued usual care (control) for 12 weeks. Spontaneous speech of every participant was recorded before, during and after both therapies. Morphosyntactic outcome measures (mean length of utterance, syntactic complexity) analysed before and after the intervention showed a significant increase. Measures of acoustic features were never studied, despite the fact that similar research with other pathological speaker groups showed change in acoustic features after following some form of speech and language therapy.

In our study, we automatically extracted 103 acoustic-phonetic features using a Praat script and the eGeMAPS script from the openSMILE toolkit. These 103 features were extracted for the recordings of 10 SLI children, 5 ‘control’ children following ‘standard therapy’ and 5 children following ‘Language Interaction Therapy’, both before and after therapy. So, in total we had 20 sets of recordings for which 20 sets of features were calculated. Two types of analyses were carried out to compare pre- vs. post-treatment data. [1] A t-test to study which features were significantly different, and what the effect sizes (Cohen’s d) of these differences were. And [2] a binary classification task called ‘Recursive Feature Elimination’ (RFE) that uses an SVM classifier. In RFE, we start with all features, and one by one the feature that contributes less to classification is eliminated. In this way, a ranking can be obtained of the features that are most relevant for binary classification. Next, with a 2nd SVM classifier, using the Matthews Correlation Coefficient (MCC) as criterion, the performance of the classification is evaluated. We’ve carried out this analysis for each individual child, and for the children together.

In this talk we will present our results, which show for which measures we observed significant changes and what the main differences are between the LIT and the control group. We then discuss these results in relation to those of previous research and consider possible future lines of research.