MetaVoice

MetaVoice is a duplex speech model designed for revenue calls. Unlike traditional voice agents that rely on half-duplex turn-taking, MetaVoice listens while it speaks, allowing for natural interruptions, overlap, and background voices. It solves the problem of over 40% of callers hanging up on voice agents within 30 seconds due to awkward pauses and failures to understand context.

Built as a single model that listens, thinks, and speaks in 350 milliseconds, MetaVoice learns conversational mechanics directly from human data. It reasons over speech—not just text—and asks for clarification when unsure. There is no need to assemble separate ASR, TTS, turn detection, or dialogue harness components.

MetaVoice is aimed at businesses running revenue calls—such as sales, customer support, and outbound campaigns—who want fewer hang-ups and higher conversion rates. It can be deployed on-premises in your own VPC, and offers guardrails, debuggability, and observability for production use.