AssemblyAI and LiveKit outline production voice agent stack
AssemblyAI has published a tutorial showing how developers can build a browser-ready voice agent on LiveKit using one AssemblyAI API key for speech-to-text, language model access and text-to-speech. The post, published on Thu, 29 Oct 2026, frames the setup as a way to reduce a multi-vendor voice stack to two main systems: LiveKit for transport and orchestration, and AssemblyAI for the models behind the conversation.
Evidence and context
According to AssemblyAI, the reference agent uses LiveKit Agents to handle WebRTC transport, room participation and turn-taking, while AssemblyAI supplies Universal-3.6 Pro Realtime for speech-to-text, the LLM Gateway for model access and assemblyai-tts-1 for synthetic voice. The company says the example can be implemented in about 100 lines of Python and supports multi-user rooms, barge-in, tool calling and turn detection without additional glue code.
The post attributes several model-level claims to AssemblyAI’s own materials. It says Universal-3.6 Pro Realtime records a 5.19% mean word error rate on AssemblyAI’s English voice-agent benchmark and a 14.4% entity error rate on names, numbers and addresses. AssemblyAI also says the model uses the agent’s own last line as input through agent_context, which LiveKit Agents 1.6.6 and later forwards automatically into the STT layer. For the voice layer, AssemblyAI says assemblyai-tts-1 streams one sentence at a time, keeps a single WebSocket session open per reply, cancels in-flight audio on barge-in without reconnecting, and publishes per-word timings so interrupted speech can be aligned to what the user actually heard.
The tutorial includes a full Python example using livekit-agents and livekit-plugins-assemblyai, along with setup steps for local console testing and deployment into LiveKit Cloud or a self-hosted livekit-server. AssemblyAI says the same API key can be used across the STT, TTS and LLM Gateway components, while the LLM Gateway speaks the OpenAI Chat Completions API so LiveKit’s OpenAI plugin can connect to it directly. The article also notes that the stack can be adapted for EU data zones by pointing the plugins to AssemblyAI’s EU endpoints.
LiveKit recently reinforced its position in the voice infrastructure market: on 2026-01-22, the company announced a $100 million funding round at a $1 billion valuation, according to company materials reported by Fortune and covered by TechCrunch and SiliconANGLE. The AssemblyAI post builds on that momentum by positioning LiveKit as the transport layer for real-time media while AssemblyAI supplies the speech and language models.
The article’s core technical claim is architectural rather than financial: it argues that transport and model inference are the two boundaries that deserve to be separate in a voice agent, and that the rest of the workflow can be handled inside one code path. The post does not supply an independent benchmark for the full end-to-end agent, and it does not add new pricing or deployment dates beyond the source material it cites.