Project Euphonia — Training Speech Recognition on Atypical Speech Through Co-Designed Data Collection

Project Euphonia is a Google Research initiative that trains automatic speech recognition (ASR) and personalised communication models on atypical speech patterns, including speech affected by ALS/motor neurone disease, dysarthria, Down syndrome and stuttering.

Rather than building a one-size-fits-all recognizer, the project aims to build personalised communication models tailored to individual speakers by recruiting people with impaired speech — via partner organisations including Team Gleason, the ALS/MND Alliance, LSVT Global and CureDuchenne — to donate speech samples and take part throughout ideation, prototyping and feedback stages.

The project expanded internationally, recruiting 132 additional speakers across four languages (38 Spanish, 14 French, 76 Japanese, 4 Hindi) in Mexico, Colombia, Peru, France, India and Japan, to test whether the recruitment-and-feedback model transfers outside English-speaking contexts.

A 2025 peer-reviewed evaluation in Frontiers in Language Sciences reports that, as of February 2025, the project had collected over 1.5 million utterances from approximately 3,000 speakers.

The paper is candid about limitations: recruitment produced a heavily imbalanced dataset across languages, formal evaluation was completed for only two of the four added languages (French and Spanish), the French evaluation carried single-rater bias risk, and the underlying datasets are not published, for privacy reasons, which limits independent replication. Google's own materials describe participant involvement as spanning ideation, design and prototype feedback, but the peer-reviewed account itself is more precise, describing 'strategic partnerships' that facilitated recruitment rather than a documented, formal co-design governance process.