On November 9, 2021, Google launched Project Relate, an Android beta app for people whose speech is hard for standard voice systems to understand - those with conditions such as ALS, cerebral palsy, Parkinson’s disease, Down syndrome, or the aftereffects of a stroke. Mainstream speech recognition is trained largely on typical speech, so it often fails the people who might benefit from it most. Project Relate addresses this by having each user record a set of phrases, which Google uses to train a model personalized to that individual’s speech patterns.
The app offers three features. Listen transcribes the user’s speech to text in real time so it can be copied into other apps. Repeat restates what the user said in a clear synthesized voice, useful for short face-to-face exchanges like ordering coffee. Assistant lets the user speak to the Google Assistant from within the app. Google said the work built on research drawing on more than a million speech samples contributed by volunteers, the same effort behind its earlier Project Euphonia.
Project Relate turned that research into a usable product, shifting from collecting data about atypical speech to giving individuals a tool that adapts to them. By training a model per person rather than expecting one model to cover everyone, it acknowledged that “atypical” speech is highly individual and that personalization, not just more data, is the path to inclusion.
Why business readers should care: Project Relate shows the power of personalized models for the long tail of users a one-size-fits-all system leaves behind, a pattern relevant anywhere a general model underperforms for a specific population.