Hi I’ve worked on similar high-speed outbound setups, and your pain points (latency, missed intent, silence gaps) are exactly what hurts conversion. For Romanian, I’ve seen better real-time results with Deepgram for low latency, or OpenAI Whisper for higher accuracy (but slower). On the voice side, tuning turn-detection, barge-in, and streaming in Retell AI is critical to avoid 2–3s dead air. Most issues usually come from overlapping speech and weak intent recovery, not just the prompt. What’s your current average round-trip latency (STT → LLM → TTS), and are you using true streaming or batch responses? @mrheisenbergw