A research group at Switzerland's ZHAW School of Engineering is tackling a gap in natural language processing (NLP) that most AI developers ignore: how to build language technologies that work with imperfect, non-standard speech. The Language AI Group, formerly known as the Natural Language Processing Group, is developing AI systems specifically designed to understand and respond to Swiss German speech from people with speech disorders, opening doors for applications like therapy chatbots and disease detection tools.

What's the Challenge With Speech Disorders and AI?

Most commercial NLP systems are trained on clear, standard speech patterns. When someone has a speech disorder caused by conditions like Parkinson's disease, stroke recovery, or developmental speech delays, current AI often struggles to understand them. This creates a real-world problem: people who could benefit most from conversational AI are often locked out of using it. The ZHAW team recognized this gap and launched a project called SpeeDi (Swiss German Speech Processing for Persons with Speech Disorders) to build language technologies that can handle the variability and imperfections in speech from people with these conditions.

The implications are significant. Imagine a child in speech therapy being able to practice conversations with an AI chatbot that actually understands their speech patterns, or a patient with Parkinson's disease getting early diagnostic feedback from a tool trained to recognize the subtle speech changes associated with the disease. These aren't hypothetical applications; they're what the ZHAW team is actively developing.

How Is the Team Approaching This Problem?

The Language AI Group combines foundational research with real-world applications, blending methods from linguistics, NLP, and artificial intelligence to enable natural communication between humans and machines. Their work spans several interconnected areas that all feed into the larger mission of making AI accessible to people with speech differences.

• Speech-to-Text Systems: The team develops speech recognition technology specifically tuned for Swiss German, including variants that handle imperfect or atypical speech patterns that standard systems miss.
• Dialogue Systems and Chatbots: They build conversational AI that can engage in meaningful back-and-forth exchanges, adapting to individual speech patterns rather than expecting users to adapt to the AI.
• Speaker Diarization: This technology identifies who is speaking in a conversation, useful for therapy sessions or clinical assessments where multiple voices are present.
• Text Classification and Sentiment Analysis: The team applies these techniques to understand not just what someone says, but the emotional tone and intent behind their words.
• Large Language Model Evaluation: As large language models (LLMs) become more common, the group evaluates how well these general-purpose AI systems perform on specialized tasks like understanding speech disorders.

The SpeeDi project, which runs from June 2026 through May 2029, represents the team's most direct effort in this space. It's not just research for research's sake; the goal is to create AI applications that people with speech disorders can actually use in their daily lives.

Why Does This Matter Beyond Speech Therapy?

The work touches on a broader principle in AI ethics and accessibility. Most technology development assumes a "standard user" with typical abilities. When AI systems are built this way, they inadvertently exclude people whose needs differ from the norm. By focusing on Swiss German speech disorders, the ZHAW team is demonstrating that NLP can be adapted to serve populations that commercial AI often overlooks.

The research also has implications for how we think about AI training data. Standard speech recognition systems are trained on thousands of hours of clear, standard speech. To build systems that work for people with speech disorders, researchers need different training data, different evaluation metrics, and different design philosophies. This requires expertise in both AI and speech pathology, which is why the ZHAW team's interdisciplinary approach matters.

"We combine foundational research with industrial applications to build new and innovative products and services, while at the same time exploring the necessary ethical and social boundaries," stated Professor Dr. Mark Cieliebak, who leads the Language AI Group.

Professor Dr. Mark Cieliebak, Language AI Group, ZHAW School of Engineering

What Other Projects Is the Team Working On?

While SpeeDi is the most directly relevant to speech disorders, the Language AI Group is pursuing several complementary projects that expand the scope of accessible NLP. One project called Multikom XR explores how AI can analyze professional communication in virtual and augmented reality environments, capturing not just words but also body language and tone. Another initiative, Meaning@Work, is developing a digital platform that helps employees find purpose and direction in their careers, using NLP to understand what matters to workers.

The team also participates in broader community-building efforts. A project called NLP Community Building (ComBi) aims to connect researchers, industry professionals, and government officials across Switzerland who work in NLP, fostering collaboration and knowledge-sharing in the field.

What ties these projects together is a commitment to making language AI more inclusive, more ethical, and more grounded in real human needs. The SpeeDi project exemplifies this philosophy by asking a simple but powerful question: if AI can understand standard speech, why can't it understand the speech of people with disorders? The answer, the ZHAW team is demonstrating, is that it can, if we design it to.

" }