The individual pieces of voice AI — transcription, reasoning, synthesis — each got fast. The 2026 achievement is assembling them into a full loop that stays under a conversational latency budget: hear, think, and speak back in well under a second.
The end-to-end challenge
A voice turn is a relay: speech-to-text transcribes the user, a language model decides what to say, text-to-speech renders the reply. Run these in sequence and the latencies stack up past the point of natural conversation. The engineering win is running them overlapped — the model starts planning on partial transcripts, synthesis begins before the full answer is generated — to hit sub-800ms voice-to-voice, and ideally the sub-500ms zone that feels truly natural.
The trick isn't making each stage fast. It's never letting a stage wait for the previous one to fully finish.
Where the time goes
Every millisecond is contested: streaming transcription to avoid waiting for the user to stop, a model fast enough to begin responding immediately, synthesis that plays its first audio in ~150ms, and a network path that doesn't add its own delay. Shaving latency is a systems problem across the whole chain, not a single model choice.
Why it's a turning point
Sub-second voice-to-voice is what turns voice agents from impressive demos into things people will actually talk to — support, assistants, real-time interpreters, accessibility tools. The capability was there; closing the latency loop is what makes it usable. This is the frontier where voice AI becomes ordinary.
