Most voice agents are three products in a trench coat: a speech-to-text service, a language model, and a text-to-speech engine, chained together. Each hop adds latency, and the total is what the caller experiences as the pause after they stop talking. It is the difference between a conversation and a walkie-talkie.
Closr's voice agent runs on a streaming speech-to-speech model instead. Audio goes in, audio comes out, and the model never round-trips through a transcript it has to wait for. That single decision is responsible for most of what makes the agent feel like a person picking up the phone.
Interruption is the hard part
People interrupt. They start answering before the question finishes, they say 'yep' in the middle of a sentence, and they expect the other side to stop talking. An agent that can't be interrupted feels broken within about fifteen seconds.
Handling that properly means being aggressive about detecting speech and conservative about what counts as speech — because the agent's own voice, coming back through a laptop speaker, looks exactly like a caller talking over it. We spent more time on the echo gate and voice-activity thresholds than on any other part of the pipeline.
An agent that can do things
Answering questions is table stakes. The agent is useful because it can act during the call:
- Look up your knowledge base mid-sentence, so the answer is grounded in a document you approved rather than the model's general knowledge.
- Check real availability and book a meeting before the call ends.
- Escalate to a human when the conversation goes past the guardrails, with hand-over wording decided server-side instead of improvised.
Knowing where it fell short
Every call is stored with its transcript, its audio, and an analysis pass that records sentiment, escalation reasons, and — most usefully — the questions the agent could not answer. Those knowledge gaps roll up across calls, so instead of listening to forty recordings you get a list of things to add to the knowledge base. One edit fixes a recurring miss.
The measure of a voice agent isn't how it sounds on the demo call. It's what the transcripts look like after a thousand of them.
That is why containment, p95 latency, and CSAT sit on the same page as the call list. The agent is a system you tune, not a switch you flip.