Presentation Information
[B-6-56]A Study on Guardrail Methods for Voice Conversational AI
〇Yoshihiko Ito1 (1. NTT, Inc.)
Keywords:
guardrail,voice-based conversational AI
With the advancement of large language models (LLMs), voice-based conversational AI has expanded rapidly. Guardrail mechanisms are essential to mitigate risks such as inappropriate speech and sensitive information handling, yet most approaches focus on text and overlook voice-specific challenges. This paper investigates guardrail mechanisms for voice conversational AI. The proposed system processes audio in real time using relay and detection modules. The relay buffers and forwards audio, while the detection module performs speech recognition and evaluates appropriateness. Since detection introduces latency, balancing responsiveness and reliable control is critical. We compare a fixed-delay approach with a dynamic-delay approach that adjusts latency based on detection results. The dynamic approach triggers additional processing only when a preliminary check detects potential issues, reducing latency in normal cases. Results show that the fixed-delay approach ensures reliable control but reduces responsiveness, whereas the dynamic approach maintains responsiveness but is affected by missed detections. These results suggest that the dynamic-delay approach is effective when sufficient detection accuracy is ensured.