I’m currently running a cold calling campaign using a custom AI agent (Andreea - Medical Clinic Campaign), but I’m encountering significant performance issues that are affecting the conversion rate. I would appreciate some guidance on how to optimize the following: 1. Latency & Long Pauses: The agent has very long response times. In a cold calling scenario, a 2-3 second delay feels like an eternity and causes the prospect to hang up or think the call is dropped. I need to reduce the "turn-taking" latency significantly. 2. Speech-to-Text (STT) & Understanding: The agent seems to "get lost" mid-conversation. Even with a clear prompt, it struggles to understand the user’s intent if they speak over the agent or use common Romanian fillers. It often fails to acknowledge the user's response and stays silent or repeats the previous logic. 3. Silence Handling: I’ve implemented a silence protocol, but the agent doesn't always trigger the "Checking 1, 2" logic correctly when the line is quiet, or it waits too long before executing the hang-up command. Current Stack/Settings: Use Case: Cold Calling (High-speed outbound) Language: Romanian Tone: Professional/Warm (1.1x speed) Goal: Lead qualification for free medical analysis via Health Insurance (CAS). What I've tried: Optimizing the prompt for brevity. Using fillers (backchanneling) to mask latency. Has anyone dealt with similar issues for the Romanian language? Are there specific LLM models or STT providers (Deepgram, Whisper, etc.) that you recommend for faster processing in Eastern European languages?