Natural Language Processing to Identify Speech and Language Characteristics in People with Autism
Natural Language Processing to Identify Speech and Language Characteristics in People with Autism
This research investigates linguistic characteristics of adults with autism spectrum disorder (ASD) using computer-assisted audio and transcript analysis, with a particular focus on women, who are often underdiagnosed due to less obvious symptoms. The aim is to provide doctors and diagnosticians with evidence-based tools that enable a more precise and gender-sensitive diagnosis of ASD and, in particular, improve the detection of autism in women.
| PhD students: | Nadine Probol |
| Supervisors: | Prof. Dr Margot Mieskes |
| Participating institutions: |
Autism spectrum disorder (ASD) was first described in the 1940s by Kanner and Asperger. Kanner identified autism as a disorder characterised by social withdrawal and communication difficulties, whilst Asperger observed that, although his patients developed language skills at an early age, they lacked emotional responsiveness. Over time, classification systems have evolved, with the DSM and ICD being the most important diagnostic tools. The ICD-10 distinguishes between early-onset autism (Kanner’s syndrome), atypical autism and Asperger’s syndrome, whilst the more recent ICD-11 groups them under the broader term ASD. The latest trends show an increase in ASD diagnoses; whilst reports vary, they generally suggest that around one in 36 children is diagnosed with autism. The prevalence is significantly higher in men than in women, although the gap is narrowing, as women are often diagnosed later in life due to less obvious symptoms. This discrepancy in diagnosis highlights the need for female patients to be better identified and supported.
This doctoral thesis focuses on research involving adults with Asperger’s syndrome and people with ASD who have high functioning and low support needs, examining both the audio data itself and its transcriptions. Furthermore, the research focuses not only on the differences between non-autistic (neurotypical; NT) and autistic people, but also on women on the autism spectrum and the differences compared to men on the autism spectrum.
Research questions
Data analysis
Where can one find usable data from people with ASD?
Is there a significant difference in the data – both in the audio and in the transcription – between men and women with ASD?
Is there a significant difference in the data – in both the audio and the transcription – between neurotypical (NT) people and people with ASD?
Characteristics of autistic speech
Can we identify a difference at the word and/or sentence level between people with ASD and neurotypical individuals, as well as between male and female people with ASD?
Can we gain more insights through the use of deep learning and transformer-based approaches than through the use of traditional machine learning methods?
Models
Is it better to use the transcripts or the audio data?
Can the results be improved by combining features from transcripts and audio data?
Do we need gender-specific models?
Analysis of results
Is there a difference between the markers used by human diagnosticians and those used by NLP-based models?
How can therapists and diagnosticians benefit from my findings?
Methodology
Data collection
Searching for openly autistic content creators on social media to ask them to make their data available for research. Automatically generating transcripts of the data collected in this way, complete with timestamps. Supplementing the data with additional descriptive tags such as singing, stammering or echolalia.
Setting up a Prolific study to enable the collection of a larger and more diverse dataset.
→ Including additional information such as comorbidities, age and diagnosis.
→ Collection of both spontaneous speech and pre-written texts.
Data analysis
Calculating statistical information directly from the data, such as pitch, mean word duration and speech rate
Qualitative analysis
Examination of speech patterns and word choice
Training models
- Training simple machine learning (ML) algorithms and deep learning (DL)/transformer models to determine whether the people speaking are on the autism spectrum or not. The focus is either solely on audio data or solely on transcription data. An ablation process is carried out to identify the best features for the models.
- Training the same algorithms and models on combined features derived from audio and transcription data. An ablation process is carried out to identify the best combination of features on which to train the models.
- Training gender-specific models to determine whether this improves the models’ performance. If performance improves, it must be assessed whether the difference is significant.
Practical application
Analysis of whether the aforementioned features can be used by doctors in diagnostic situations to improve their diagnostic process.
Steps taken so far
Following an extensive literature review, significant gaps in the research were identified. As part of the data collection from social media, influencers within the relevant spectrum were specifically contacted and asked for permission to use their public data for research purposes. This systematic approach resulted in a comprehensive dataset comprising more than 282 hours of data from women and more than 107 hours of data from men. Initial findings have already been published.
In parallel, a Prolific study was set up to collect more diverse data in order to broaden the research base. The corresponding survey is currently being fine-tuned and will be published shortly, once funding has been finalised. In addition, contacts have been made with a medical expert to strengthen the project’s interdisciplinary focus.
Objective
The aim is to provide doctors and diagnosticians with evidence-based tools that enable a more precise and gender-sensitive diagnosis of ASD and, in particular, improve the detection of autism in women.
References
Hans Asperger. 1944. ‘Autistic psychopaths’ in childhood. Archiv für Psychiatrie und Nervenkrankheiten, 117(1):76–136.
Leo Kanner et al. 1943. Autistic disturbances of affective contact. *Nervous Child*, 2(3):217–250.
Nadine Probol and Margot Mieskes. 2024. Autism detection in speech – a survey. In Proceedings of EACL 2024, 1115–1125.
Nadine Probol and Margot Mieskes. 2025. Female Autism in Natural Language – A Corpus Paper. In Workshop Proceedings of the 19th international AAAI Conference on Web and Social Media (IWCSM 2025).
