I built an AI English speaking coach — what was technically hard

I spent the past year building Elispeak , an AI English speaking coach. The user-facing pitch is simple — talk to an AI tutor, get instant pronunciation and fluency feedback, practice TOEFL / IELTS / CELPIP speaking tasks on demand. Under the hood, a few things turned out to be much harder than I expected. This is a note to myself, and to anyone else building voice-first language tools. 1. Real-time ASR + scoring latency is the whole product The promise of "instant feedback" falls apart at 4 sec