This paper examines two guardrail strategies for speech-to-speech LLM assistants in automotive applications: transcript-based checks and tool-based checks. Its empirical evaluation reports that both approaches are generally insufficient for industrial deployment. Even computationally inexpensive checks can add 0 to 1.4 seconds of latency to each response, while tool-based safeguards may introduce non-deterministic tool-call behavior. The authors use these findings to identify open challenges for deploying programmable safety controls in end-to-end, natural-sounding in-car assistants.
No heat snapshots are available in the last 24 hours.