To be clear, I haven't shared this with anyone because I'm not yet sure that the content is useful and correct.
As far as where I'm at:
- I've listened to it in my target language for N hours. To my ear, it sounds correct and I've learned some new words that I then heard used consistently in native media.
- Next, I'd like to set it to teach me a language that I already know, so that I can more reliably and easily spot errors. This will require some changes, since my target language is currently hardcoded.
- Longer-term, validation based on languages I speak can't generalize 100% to other languages, nor can validation of version N make assertions about version N + 1. Correctness would benefit from native speakers periodically checking results, and usefulness would benefit from user feedback (even if only in the form of engagement or lack thereof).
I first generate a script to be handed to a TTS model. For this step, Claude 3.5 Sonnet works well. For voice synthesis, I've been using Google Cloud's Text-to-Speech API and it's been adequate.
As far as where I'm at: - I've listened to it in my target language for N hours. To my ear, it sounds correct and I've learned some new words that I then heard used consistently in native media. - Next, I'd like to set it to teach me a language that I already know, so that I can more reliably and easily spot errors. This will require some changes, since my target language is currently hardcoded. - Longer-term, validation based on languages I speak can't generalize 100% to other languages, nor can validation of version N make assertions about version N + 1. Correctness would benefit from native speakers periodically checking results, and usefulness would benefit from user feedback (even if only in the form of engagement or lack thereof).