The doctoral dissertation in the field of General Linguistics will be examined at the Philosophical Faculty at Joensuu campus.
What is the topic of your doctoral research? Why is it important to study the topic?
My doctoral research focuses on the digitalisation of speech and language therapy (SLT) through automatic speech recognition (ASR) and natural language processing (NLP), with a case study on German-speaking people with aphasia. The goal is to develop methods that enable automated analysis of spoken language and provide meaningful, personalised feedback in therapy exercises. The methods should also be scalable to other languages and use cases.
This topic is important because effective SLT requires frequent and intensive practice, yet access to professional therapy is often limited. Digital solutions can support independent training at home, but current tools rarely provide detailed feedback on speech production. By integrating linguistic analysis with AI technologies, this research contributes to making therapy more accessible, scalable, and effective, ultimately improving rehabilitation outcomes and quality of life for individuals with language disorders.
What are the key findings or observations of your doctoral research?
The key outcome of this research is the development of a modular linguistic analysis pipeline that processes spoken input at multiple levels – phonetic, lexical-semantic, and grammatical – and enables automated, detailed feedback in SLT exercises or similar contexts. The system integrates open-source ASR and NLP tools and adapts them to atypical speech. The adaptation is knowledge-based and does not require an extensive amount of sensitive clinical data or computational resources.
Carefully selected open-source ASR models can outperform commercial solutions in recognising atypical speech and reach a level acceptable for implementation in SLT applications. The dissertation also describes methods for improving recognition.
To my knowledge, my thesis is the first study that introduces semantic analysis of users’ errors using lexical networks. It also elaborates on a method for inferring users’ intended answers when the ASR output does not match an existing word.
Further novelty lies in combining these components into a transparent and scalable framework that goes beyond simple right/wrong feedback. The approach is validated on datasets collected from German-speaking people with aphasia (PWA), including speakers of certain German dialects, and the ASR component is also validated on German-speaking multilingual children.
The work demonstrates that such solutions are feasible even under real-world constraints, making them highly relevant for both research and practical applications.
How can the results of your doctoral research be utilised in practice?
The results of my doctoral research can be applied in the development of digital speech and language therapy applications that support independent practice outside clinical settings. The proposed pipeline enables automated feedback in oral exercises, helping users improve pronunciation (to a certain extent), vocabulary, and grammar without constant supervision from a therapist. In practice, this can increase therapy intensity and accessibility, especially for individuals with limited access to professional care. The approach can also support clinicians by complementing in-person therapy and reducing routine workload.
Beyond aphasia rehabilitation, the methods are transferable to other contexts, including therapy for different language disorders, language learning, and assessment tools. This makes the research broadly applicable in digital health and educational technologies.
What are the key research methods and materials used in your doctoral research?
My doctoral research combines methods from applied linguistics, phonetics, and computational linguistics. The core approach is the development and evaluation of a linguistic analysis pipeline that integrates ASR and NLP tools. Key steps included: evaluating multiple ASR systems on typical and atypical speech data, selecting and adapting robust open-source ASR models, selecting suitable NLP resources, developing modules for phonemic, semantic, and grammatical analysis, implementing error classification and feedback mechanisms, testing the system in the context of speech therapy tasks (e.g. picture naming). The research used experimental datasets of atypical speech, including aphasic speakers of the target dialect group. Both quantitative evaluation (e.g. recognition accuracy) and qualitative analysis of errors were applied.
The methods were developed in collaboration with clinical and technical stakeholders.
Is there something else about your doctoral dissertation you would like to share in the press release?
The work was conducted within the aphaDIGITAL project (https://inno-tdg.de/projekte/aphadigital/). The project was sponsored by by the Federal Ministry of Research, Technology and Space of Germany (formerly Federal Ministry of Education and Research).
Further information:
Eugenia Rykova, [email protected]