Similarity Metrics for Aligning Children's Articulation Data

I. Background This paper concerns the implementation and testing of similarity metrics for the alignment of phonetic segments in transcriptions of children's (mis)articulations with the adult model. This has an obvious application in the development of software to assist speech and language clinicians to assess clients and to plan therapy. This paper will give some of the background to this general problem, but will focus on the computational and linguistic aspect of the alignment problem. 1.1. Articulation testing It is well known that a child's acquisition of phonology is gradual, and can be charted according to the appearance of phonetic distinctions (e.g. stops vs. fricatives), the disappearance of childish lnispronunciations, especially due to assimilation ([gtN] for dog), and the ability to articulate particular phonetic configurations (e.g. consonant clusters). Whether screening whole populations of children, or assessing individual referrals, the articulation test is an important tool for the speech clinician. A child's articulatory development is usually described with reference to an adult model, and in terms of deviations from it: a number of phonological "processes" can be identified, and their significance with respect to the chronological age of the child assessed. Often processes interact, e.g. when spoon is pronounced [mun] we have consonant-cluster reduction and assimilation. The problem for this paper is to align the segments in the transcription of the child's articulation with the target model pronunciation. The task is complicated by the need to identify cases of "metathesis", where the corresponding sounds have been reordered (e.g. remember --+ [mtremb~]) and "merges", a special case of consonant-cluster reduction where the resulting segment has some of the features of both elements in the original cluster (e.g. sleep , [tip]). It would be appropriate here to review the software currently available to speech clinicians, but lack of space prevents us fi'om doing so (see Somers, forthcoming). Suffice it to say that software does exist, but is mainly for grammatical and lexical analysis. Of the tiny number of programs which specifically address the problem of articulation testing, none, as far as one can tell, involve automatic alignment of the data.