Interestingly enough, I was working on something very similar this past few of days[1]! A couple of things that I did differently -
One, in the post-transcription cleanup phase, instead of programmatically fixing the text, I just feed the text with a prompt + dictionary of words into an LLM. Initially tried doing this with the on-device model from Apple in macOS 27, but that was too slow. I've now been playing around with gpt-oss-20b on Groq, which appears to be fast and effective. Although I'm curious how well your programmatic approach works, I might try some sort of a bake off over the next few days. But in the long run, I think I might still stick with an LLM-based formatter because you can have far more interesting formatting considerations or formatting prompts based on the app that you are using or the style or tone of your dictated text.
Two - When building the dictionary of words to fix, I'm also including a recording of the word and having the transcription model just predict draft out variants to use as replacements.
One, in the post-transcription cleanup phase, instead of programmatically fixing the text, I just feed the text with a prompt + dictionary of words into an LLM. Initially tried doing this with the on-device model from Apple in macOS 27, but that was too slow. I've now been playing around with gpt-oss-20b on Groq, which appears to be fast and effective. Although I'm curious how well your programmatic approach works, I might try some sort of a bake off over the next few days. But in the long run, I think I might still stick with an LLM-based formatter because you can have far more interesting formatting considerations or formatting prompts based on the app that you are using or the style or tone of your dictated text.
Two - When building the dictionary of words to fix, I'm also including a recording of the word and having the transcription model just predict draft out variants to use as replacements.
[1] https://github.com/goose42/TalkToMyMac