For years, voice AI meant three round-trips (speech-to-text, LLM, text-to-speech), noticeable latency and a distinctly robotic feel. The current generation of realtime speech-to-speech models has quietly closed most of that gap. Our voice deployments now hit ~400ms first-response latency with natural prosody, interruptions and back-channelling.
The engineering challenges are also very different from a chatbot's.
Interruption handling is table stakes
People interrupt. They correct themselves mid-sentence. They mumble. A voice agent that requires clean turns feels broken within thirty seconds. Every realtime session we ship uses server-side voice activity detection and cooperative cancellation: the moment the user starts speaking, the current model utterance is cut off, and the model is given the partial audio as context.
The prompt is only half the personality
Voice, pace and pauses do more work than the words. We spend real time on voice selection, target speaking rate and the model's "listening posture" — small acknowledgements while the user speaks. Prompt engineering here looks a lot more like directing an actor than writing a brief.
Where voice-first agents genuinely shine
- Field workers who can't type — logistics drivers, technicians, healthcare staff.
- Outbound qualification calls where a human tone dramatically improves conversion.
- Multilingual support in markets where written literacy is uneven.
- Accessibility surfaces where voice is the primary modality by design.
Want to talk through this on your project?
We work with teams shipping AI, automation and intelligent products end-to-end. If any of this maps to what you're building, we'd love to compare notes.
Start a conversation