Towards Automatic Detection of Developmental Language Disorders

K.m. Knill, Abhi Sharma, Charles McGhee, Elizabeth Worster, Rao Ma, Harriet J. Smith, Matthew H. Davis, Penny Karanasou2025 · DOI

Abstract

Developmental Language Disorders (DLD) affect over 7% of children in primary classrooms and have lifelong consequences. DLD is typically under-diagnosed due to difficulties in administering and scoring standardized tests that assess phonological abilities. People with DLD or related impairments, such as dyslexia, have well-recognized issues with phonological processing. This study utilizes a new dataset from adult speakers—21 self-declared dyslexics and 30 controls—focusing on two speech elicitation tasks: Non-word Repetition (NWR) and Rapid Automatized Naming (RAN). We established a pipeline of automatic approaches to detect errors at the phone and word level and to attempt speaker classification. Phone-level automatic speech recognition (ASR) and articulatory feature alignment achieved reasonable detection of mispronunciations (F1 0.826). However, classification results were limited by the small dataset size. We plan to apply this pipeline to larger datasets and other diagnostic tasks.

Introduction

Developmental Language Disorders (DLD) represent a significant public health challenge, affecting approximately 7% of children. These disorders often co-occur with dyslexia and lead to long-term academic and social difficulties. Despite their prevalence, DLD remains frequently under-diagnosed. The current gold standard for diagnosis involves manual administration and scoring of tests like Non-word Repetition (NWR), where subjects repeat phonologically complex made-up words (e.g., "vudanuwub"). This manual process is time-consuming, subjective, and often fails to capture the granular nature of phonological errors. This research explores automated speech processing tools to provide a scalable, objective baseline for clinical diagnostic tasks.

Methodology

Our research established a multi-pronged pipeline to analyze speech data, focusing on three distinct computational approaches to capture phonological deviations.

Overview of the three-pronged automated analysis pipeline including phone recognition, articulatory alignment, and embedding-based classification.

1. Phone Recognition-based ASR

Traditional ASR models often suffer from a "correction trap" when processing non-words, as they attempt to map unfamiliar sounds to a known lexicon. We compared the performance of Whisper-large-v2 against a phone-level model, WavLM. By operating at the phoneme level, WavLM avoids lexical bias, which is essential for clinical tasks where the target stimuli do not exist in standard dictionaries.

2. Articulatory Feature Alignment

This novel approach maps audio signals to continuous articulatory features representing vocal tract movements (e.g., voicing, nasality). We utilized Dynamic Time Warping (DTW) to calculate an "alignment cost" between the target stimulus and the participant's response. This method allows for the detection of sub-phonemic variations that traditional categorical scoring might overlook.

3. Embedding-based Classification

Finally, we employed self-supervised learning (SSL) representations from W2V2 and Whisper to train linear classifiers. The goal was to determine if high-dimensional speech embeddings could distinguish between dyslexic and neurotypical speakers based on their overall speech patterns during the NWR and RAN tasks.

Results and Discussion

Error Detection and ASR Performance

The articulatory feature alignment method (Art + VVN) proved highly effective for detecting mispronounced non-words. It achieved an F1 score of 0.826 and a Recall of 0.929, demonstrating its utility as an automated scoring tool.

Scatter plot showing the distribution of normalized alignment errors across non-word stimuli for correct and incorrect productions.

As shown in the figure above, the alignment cost effectively separates correct from incorrect productions across various non-word IDs. Furthermore, phone-level ASR (WavLM) significantly outperformed word-level models, reducing the Phone Error Rate (PER) from approximately 59% to 31% for the control group.

Classification Challenges

Despite the success in error detection, classifying individual speakers as "Dyslexic" versus "Control" was largely unsuccessful. The best-performing model reached only 58% accuracy, which is barely above chance. This is likely due to the small sample size (N=52) and the use of adult participants. Adults with language disorders often develop strong compensatory strategies, making their speech patterns more similar to neurotypical controls than those of children with active DLD symptoms.

Conclusion

This study demonstrates that automated pipelines, particularly those utilizing articulatory feature alignment, can effectively detect phonological errors in diagnostic speech tasks. While speaker classification remains a challenge with small adult datasets, the high F1 scores for mispronunciation detection suggest these tools could significantly reduce the burden on speech-language pathologists. Future work will focus on applying this pipeline to larger datasets of children, where the signal for DLD is expected to be more pronounced.

Summary updated

Abstract

Developmental Language Disorders (DLD) affect over 7% of children in a primary classroom and have lifelong consequences. DLD is typically under-diagnosed due to difficulties in administering and scoring standardised tests that assess their phonological abilities. People with DLD or a related impairment (Dyslexia) have well-recognised issues with phonological processing. We here work with a new dataset from adult speakers: 21 self-declared dyslexics and 30 controls. We focused on two speech elicitation tasks with established links to phonological processing: Non-word Repetition (NWR) and Rapid Automatized Naming (RAN). In NWR, participants hear sets of well-formed non-words (e.g. dipe, vudanuwub, length 1-6 syllables) and repeat each non-word, as accurately as possible. In RAN, participants rapidly name sequences of digits or objects presented in a visual array. We have established a pipeline of automatic approaches to detect errors at the phone and word level and to attempt to classify whether speakers have a DLD or not. Phone level automatic speech recognition with cross-phone error scoring across stimuli and participant responses yielded the lowest error rates, although these remain higher (at 20-26%) than desirable. Aligning continuous articulatory features from the participant’s speech with the stimuli achieved a reasonable detection of when words have been mispronounced (F1 0.826). Classification results were disappointing, with no distinction found between the dyslexic and neuro-typical control participants due to a limitation of the small data set. We plan to apply this pipeline to larger data sets and other diagnostic tasks.
BibTeX
@misc{knill2025towards,
  author = {K.M. Knill and Abhi Sharma and Charles McGhee and Elizabeth Worster and Rao Ma and Harriet J. Smith and Matthew H. Davis and Penny Karanasou},
  title = {{Towards Automatic Detection of Developmental Language Disorders}},
  year = {2025},
  doi = {10.33774/coe-2025-3nx1h},
}
Figure