Vellum adds Voice Mode to personal AI assistant

Vellum is routing voice through its existing agent loop, so spoken requests can use the same tools, memory, approvals, and identity as text chat.

The system transcribes speech with Deepgram, processes it through the assistant, and speaks replies through ElevenLabs. Vellum says it kept a cascade setup because native speech-to-speech models would not expose the agent loop it wanted. The piece highlights custom handling for pauses, interruptions, long tool calls, and audio-only progress updates. Voice continuity is meant to carry across devices, with longer tasks moving back to text when needed. TestingCatalog's note