Can you run it in reverse for speech recognition?

We will release an STT model as well.

no, but whisper has a 39M model: https://github.com/openai/whisper