May I ask who’s calling? How Decagon handles the challenge of voice authentication | Decagon
12/4/2025
This post details the technical challenges and solutions for voice authentication in AI agents. It breaks down why voice authentication is difficult (transcription of identifiers, multitasking callers, unreliable phone numbers), presents two primary authentication models (pre-authenticated and mid-conversation) with their respective implementation strategies and tradeoffs, and discusses the selection of appropriate information identifiers based on ease of retrieval, security, and reliable voice capture. It also introduces the concept of incremental agent authentication, where agent privileges are granted based on evolving user identity assurance, moving away from universally privileged tokens. Finally, it highlights Decagon's investment in fine-tuning Speech-to-Text (STT) models for authentication-centric utterances to improve precision and reliability.