Our team is developing a text-to-speech model that is small and efficient enough to run directly on everyday low-power devices, so the voice can work on the device itself, without sending text to the cloud.
Making a model smaller and faster can change how natural it sounds. Computers can estimate speech quality, but the most reliable judge is still the human ear. That is why we ask people to listen and rate how natural different versions of the voice sound.
Your ratings show us which design choices keep the voice natural while making it efficient. More details will be shared once the work is published.
