Created this AI accent/dialect coach in Vercel x Deepmind hackathon and won 2nd place!
It uses wav2vec2 to fit to phonemes and measure how far off you are from the reference pronunciation.
This allows you to analyze at pronunciation-symbol (phoneme) level not at the english level (so because of this llms can't tell you how well you did in detail)

opensourced: https://github.com/chattermint/chattermint

This is an awesome project and it's amazing what these models are capable of now.