Currently people with aphasia can only type very few words per minute. Famously, Stephen Hawking could only type less than 10 wpm. Predictions are a big part of increasing this speed. By using very recent Smart Reply LSTM models, we hope to increase this.
What it does
Transcribes speech from the user's microphone, uses TensorFlow Lite to generate smart reply suggestions and allows the user to read out of those suggestions.
How we built it
We build a TensorFlow Lite wrapper in C, and call it from an Express.js child process. This passes it to a React frontend.
Challenges we ran into
We were hoping that we could use the Python Interpreter for TensorFlow Lite, but unfortunately it does not have full support.