A Connectionist E valuation of S chemes to M easure the P honological S tructure of W ords

Jennifer Hayes and Peter Howell

Psychology Department

University College London

Gower St., London, WC1E 6BT

The EXPLAN model (Howell, 2004) argues that speech production involves independent planning and execution processes. Fluency failures such as repetition of prior words, pausing, prolongation and repetition of parts of the current word occur when the word to be produced is not ready (the planning process is not complete) by the time the execution of the previous word is concluded. EXPLAN can explain the behaviour of both fluent speakers and also the speech of speakers who stutter (SWS).

There are two influences that lead to the plan for the current word not being ready in time a) execution time of the prior word and b) planning time of the current word. Rapid execution of a preceding word advances the time at which the current word needs to be ready. Factors that make the current word complex lengthen its planning time and delay when it is ready. Execution time of the previous word is easily measured but an effective measure of word complexity has been harder to obtain. The connectionist models reported here compared two schemes which measure the phonetic complexity of words to see if they could predict which items in a stream of naturalistic speech from SWS were likely to be stuttered. Two multi-layer perceptron models were trained on the same data set of 6743 spontaneously produced words from eight SWS. For one model the words in the training set were represented by index of phonetic complexity (IPC) codes (Jakielski, 1998) and for the other the input was made up of analysis of phonetic structure (ANOphS) (Diment, Howell and Harris, in preparation) codes representing the words. The model was required to output a stuttered or a fluent response depending on the phonetic complexity of the word measured by either the IPC or ANOphS scheme. The models were tested using a further set of 1000 unseen words. While the model trained using the IPC codes could not learn to discriminate between stuttered and fluent items the ANOphS model demonstrated that it could learn the task. Further modelling is planned to test the EXPLAN model in which the ANOphS coding scheme will be used as a metric of phonetic complexity.

Howell, P. (2004). Assessment of some contemporary theories of stuttering that apply to spontaneous speech. Contemporary issues in Communication Science and Disorders, 31,123-140

Jakielski, K. J. (1998). Motor organization in the acquisition of consonant clusters. PhD thesis, Univeristy of Texas at Austin. Ann Arbor Michigan, UMI Dissertation services.

Diment, R., Howell, P., and Harris, J. (in preparation), Development of a procedure for the Analysis of Phonological Structure (ANOPhS) and it’s application for predicting stuttering in English speaking adults.