A conference talk from NDC London 2026 by Nathaniel Okenwa, a developer evangelist at Twilio, covering how to build voice interfaces for LLMs. The talk explores the history of voice AI from Siri and Alexa to modern LLMs, discusses the uncanny valley problem in voice interfaces, and demonstrates practical techniques including interruption handling, voice interstitials for long-running tasks, and contextual memory management. The speaker live-codes a voice-controlled presentation tool using Twilio, Deepgram, OpenAI, and ElevenLabs, and showcases Twilio's Conversation Relay product and Segment for customer memory. Key architectural patterns covered: speech-to-text → LLM → text-to-speech pipeline, state management for interruptions, and multi-channel context awareness.
Sort: