Given the size of the Sindarin team (3 AFAICT), that mostly looks like a clever combination of existing tech. There are some speech APIs that offer word-by-word

Given the size of the Sindarin team (3 AFAICT), that mostly looks like a clever combination of existing tech. There are some speech APIs that offer word-by-word realtime transcription (Google has one), assuming most of the special sauce is very well thought out pipelining between speech recognition->LLM->TTS (not to denigrate their awesome achievement, I would not be interested if I were not curious about how to reproduce their result!)