Automated Screening of Speech Development Issues in Children by Identifying Phonological Error Patterns

Abstract

A proof of concept system is developed to provide a broad assessment of speech development issues in children. It has been designed to enable non-experts to complete an initial screening of children's speech with the aim of reducing the workload on Speech Language Pathology services. The system was composed of an acoustic model trained by neural networks with split temporal context features and a constrained HMM encoded with the knowledge of Speech Language Pathologists. Results demonstrated the system was able to improve PER by 33% compared with standard HMM decoders, with a minimum PER of 19.03% achieved. Identification of Phonological Error Patterns with up to 94% accuracy was achieved despite utilizing only a small corpus of disordered speech from Australian children. These results indicate the proposed system is viable and the direction of further development are outlined in the paper.

Improving Child Speech Disorder Assessment by Incorporating Out-of-Domain Adult Speech

“…Other approaches have exploited the fact that many assessments of speech and articulation are made with known texts. The improvement in recognition this can yield for clinical ASR has been demonstrated previously by this group [17], where an improvement of 33.1% in phoneme recognition was achieved when the target text was exploited by the constrained decoder. Other groups have had similar success with this approach, improving detection of pathological voice by 20% [18].…”

“…The input to these models is speech made up of target words in isolation, elicited through a picture naming task. The acoustic model used in the previous system [17] used a hierarchical neural network (HNN) with long temporal context features over a 310ms window. This HNN cascaded a pair of neural networks (NN) into a third NN, with the third NN trained on the concatenated posterior probabilities of the first two NN.…”

“…Research has also investigated the application of ASR to children's speech development, assessing autism spectrum disorders [13], childhood apraxia of speech [14], cleft lip and palate [15] and general language development [16]. Despite the increasing abundance of such research, the focus remains primarily on therapy rather than screening, with very little work addressing the need for a broadly applicable screening tool [16,17]. A key challenge which has constrained clinical ASR is the scarcity of data available for the target population.…”

“…In this paper, we investigate whether an approach utilising transfer learning combined with previously developed methods for constrained decoding based on knowledge of the target text [17], can be applied to improve the detection of phonological errors in children's speech. A large and readily available dataset of adult Australian speech is used to train a base DNN, which is then transferred to train our target DNN of disordered Australian children's speech.…”

“…A large and readily available dataset of adult Australian speech is used to train a base DNN, which is then transferred to train our target DNN of disordered Australian children's speech. Figure 1 presents the three stage assessment system (the 'proof of concept' which was developed in [17] and [22]). This paper presents improvements to the initial stage, whilst maintaining the novel constrained HMM decoder and phonological error pattern (PEP) detection stages from previous work.…”