BIGGER PICTURE Accessibility & independence
The speech recogniser learns the speaker
When speech differs from the training data, a standard voice interface can repeatedly misunderstand its user. Project Relate reverses the burden: it builds a speech model around one person’s recordings.

- Personal training
- 500 recorded phrases in the app
- Outputs
- Text, spoken repetition and voice typing
- Offline option
- Listen works after the model is prepared
How the idea works
Pair words and voice
Known phrases provide labelled examples of personal speech.
Personalise recognition
The model learns the speaker’s recurring sound patterns.
Choose an accessible output
Recognised words become text, synthetic speech or typed input.
Being misunderstood is a system mismatch
Atypical speech can follow conditions such as cerebral palsy or stroke. A recogniser trained mainly on typical speech may perform poorly for these voices. Project Relate was introduced as a communication tool that learns individual patterns rather than expecting everyone to match the same pronunciation.
Sources: [1]
A personal set of examples
Users record phrases whose written words are known. Pairing the audio with that text gives the model examples of how this person produces speech. The current help page describes five hundred recordings, including optional custom phrases, before a personalised model is created.
Sources: [2]
One voice, several ways to be understood
The model can turn speech into readable text or repeat it through synthetic speech. Voice typing extends its use into other apps. These are practical bridges to a conversation partner or a text field. The Listen feature can operate without an internet connection once the model is available.
Sources: [2]
Personal does not mean equally successful for everyone
Underlying Google research showed large recognition improvements with personalised models, including severely impaired speech. That benchmark is not a measured outcome for every Relate user. The product itself warns that rapidly changing or very severely affected voices may remain difficult and that helpfulness varies between personal models.