Adapt4Me advances inclusion through digital tools that adapt to the user, not the other way around. Our goal is that everyone is understood despite impairment in speech.
Personalized AI Speech Recognition for People with Speech Impairments
We believe that everyone deserves to be understood. While speech recognition has become widely available, a large community is still being left behind: people with speech disabilities. We believe that individuals should not have to adapt to technology. Instead, technology must adapt to the individual. That is why we are working under high-pressure to make speech recognition personalized and accessible to everyone.
How it work
It starts by listening to you. During onboarding, the app builds a personal speech profile based on your pronunciation and style. As you use it, simple corrections help the system learn continuously, making speech recognition more natural and reliable over time.
Who it’s for
It is made for people whose voices are not well understood by standard speech technology. It also supports families, caregivers, and social environments by enabling clearer communication and reducing everyday misunderstandings.
Why it matters
It removes communication barriers by adapting to the individual, not the average. It supports independence, builds confidence, and helps people reconnect with the joy of being understood in daily conversations.
Explanatory Video adapt4me
Support tool for speech therapists
A second focus is supporting speech therapists in diagnosis and treatment. Adapt4Me includes an algorithm for automated phoneme analysis, designed for use in clinical and educational settings. This tool does not replace professional expertise; it complements it with objective, data-driven insights for assessment, monitoring, and personalized intervention. This can make speech therapy more precise, efficient, and accessible.
The Research
Semantic re-chaining
A method to personalize speech recognition with minimal data by generating sentence variations from a small set of words. This expands training data and improves recognition, even in low-data scenarios.
Variational LoRA
A lightweight adaptation technique that enables robust personalization from limited speech samples. It improves model stability and handling of uncertainty, with validated results in both English and German. See movie→
Phoneme uncertainty
An approach that identifies difficult speech sounds based on model uncertainty and prioritizes them during training. This leads to improved accuracy and aligns closely with expert speech therapy assessments.
Adapt4Me
A user-centered system that makes personalized speech recognition accessible from initial recording to continuous adaptation. It combines efficient speech sample selection, Variational LoRA, and model-based uncertainty, allowing users to improve their speech recognition with minimal effort.
Clinical Context for Speech Recognition
We investigate whether modern audio-language models can use additional clinical information to better recognize dysarthric speech. Diagnostic information and clinical descriptions provide little benefit through direct prompting, while targeted model adaptation can lead to substantial improvements and reveals where current models still reach their limits.
The Team
This project was developed by an interdisciplinary research team at ETH Zurich, bringing together strong expertise in machine learning, speech technology, and accessibility.
Roman Boehringer
Project lead for adapt4me and labmanager in the Grewe Lab
Lab manager and scientific coordinator in Prof. Benjamin Grewe’s team at ETH. Develops speech recognition algorithms for people with speech impairments, focusing on low-data and personalization.
Google-Scholar
Niclas Pokel
Doctoral candidate in the Grewe Lab
Former master’s student, now PhD candidate in Prof. Benjamin Grewe’s team. Continues developing speech recognition algorithms for people with speech impairments.
Yingqiang Gao
Postdoctoral fellow in the institute for computerlinguistics
I am currently a postdoc at the Institute for Computational Linguistics at UZH. Besides speech recognition, my work focuses on developing tools for accessibility.
Google-Scholar Website
Pehuen Moure
Doctoral candidate with Shih-Chii Liu
PhD student at the Institute of Neuroinformatics (ETH/UZH), working on sensor systems and speech recognition for impaired speech.
Google-Scholar Website
Master students contributing to the project: Philipp Guldimann, Jingjing Qiu, Fabian Bruelisauer, Bilal Bounajma and Orlando Leone.
Be Part of Our Vision
We believe that every voice deserves to be understood. Our goal is a world in which communication is accessible, personal, and inclusive, regardless of how a voice sounds.
Adapt4Me is currently under development. To further develop the tool in a targeted way and adapt it to real needs, we are seeking exchange with people and institutions interested in inclusive communication.
If you have questions, are interested in the project, like the idea, would like to support us, or are considering funding or a donation, we would be pleased to hear from you.
Our Partners
For questions related to speech and speech therapy, we work closely with Samra Hamzic, the Head of Research at the Hochschule für Logopädie Ostschweiz.
We recently received funding for our project from Stiftung Cerebral. We also work closely with Stiftung Cerebral to benefit from the expertise of its users.
