“The common production architecture for voice AI agents is a speech-to-text, text-to-text, and text-to-speech pipeline, rather than a direct voice-to-voice model, to ensure accuracy.”