@schlawg thanks for the post! This is something I've been thinking about implementing myself, but I would have had to implement it as a third party solution because I wouldn't have the skills to integrate/contribute it to Lichess this nicely.
Do you think there is a way to get the speech recognition up to a higher accuracy rate, since the field is moving so fast nowadays and it's been half a year? I was trying out OpenAI's Whisper today and it transcribes chess moves so well out of the box. I understand that you want the model people need to download to be fairly small, but maybe there could be some option somehow to download a better one if the user wants?