@Jarrah ah thanks, that's exactly the direction I was thinking. On recordings: our client runs a solar energy service in Benin, ~50k calls/month, 50% in Fon. They've been operating for years so there's a significant volume of real call audio available. Getting a clean batch shouldn't be an issue. On response time: we're targeting sub-1.5s latency end-to-end. Is Coqui XTTS realistic at that threshold, or does the fine-tuning overhead push it over? Also open to other TTS options if you've seen something faster work in practice for low-resource languages.