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

Abstract

This paper describes the continued development of a system to provide early assessment of speech development issues in children and better triaging to professional services. Whilst corpora of children's speech are increasingly available, recognition of disordered children's speech is still a data-scarce task. Transfer learning methods have been shown to be effective at leveraging out-of-domain data to improve ASR performance in similar data-scarce applications. This paper combines transfer learning, with previously developed methods for constrained decoding based on expert speech pathology knowledge and knowledge of the target text. Results of this study show that transfer learning with out-of-domain adult speech can improve phoneme recognition for disordered children's speech. Specifically, a Deep Neural Network (DNN) trained on adult speech and finetuned on a corpus of disordered children's speech reduced the phoneme error rate (PER) of a DNN trained on a children's corpus from 16.3% to 14.2%. Furthermore, this fine-tuned DNN also improved the performance of a Hierarchal Neural Network based acoustic model previously used by the system with a PER of 19.3%. We close with a discussion of our planned future developments of the system.

Untranscribed Web Audio for Low Resource Speech Recognition

“…Domain adaptation is a well studied problem in automatic speech recognition. Multiple techniques have been developed for the adaptation of acoustic models, such as transfer learning [1,2] and feature mapping [3,4]. Semi-supervised and lightly supervised adaptation techniques use a base model trained on out-of-domain supervised data to generate targets on in-domain unsupervised data.…”

UltraSuite: A Repository of Ultrasound and Acoustic Data from Child Speech Therapy Sessions

“…We are in the process of collecting further data from 120 children with SSD in the Ultrax2020 project following the protocol described in [36]. In addition, we intend to add other available data to our repository, including adult data and alternative forms of articulatory imaging techniques (e.g., MRI of vocal tracts), all of which can be used in data augmentation methods [17,15,10]. We encourage other researchers to contribute by submitting their data for us to standardise and add to this repository.…”

“…Machine learning has the potential to automate much of this work, leading to better outcomes for patients without increasing workload for pathologists, but publicly available data that could facilitate this work is scarce. Existing work reports results on adult data [7,8,9], data that is not publicly available [10], or data that is in proprietary format [11,12,13]. Additionally, child speech processing and disordered speech processing are both known to present many challenges [14,15,16,17].…”

Multidisciplinary Perspectives on Automatic Analysis of Children’s Language Samples: Where Do We Go from Here?

“…Much of the recent research on ASR for children has been focused on how data on adult speech can be used during training to improve recognition for children. Authors of [48, 49] investigated how to fine-tune models trained on adult speech recognition with child data. Fine-tuning is the process of first training a machine learning model on one domain with large amounts of data (in this case adult speech) and then retraining either parts of the model or the whole model on the target domain (here children’s speech).…”