Voice agent testing is software that tests an AI voice agent by having imaginary callers, with set personas, hold whole conversations with it, and then scoring each conversation. A voice agent is software that answers phone calls by itself, using a language model, the kind of AI behind chatbots, to understand the caller and decide what to say. A persona is a description of an imaginary caller: who the caller is, what they want and how they talk.
Why does it need a new kind of test? An ordinary test compares an output with the exact answer it expects. A voice agent built on a language model never answers the same way twice. Two correct answers may use completely different words. So an exact-match test fails even when the agent is right. The page on testing non-deterministic software explains this problem. The test has to judge the whole conversation instead. Was the caller's problem solved? Were the facts correct? Was anything said that should not have been said?
Here is how a test run works. The team writes scenarios, such as "renew a lapsed policy" or "complain about a wrong deduction". For each scenario, the tool creates several personas. One talks fast in Hinglish. One interrupts the agent mid-sentence. One first gives a wrong policy number. One is angry. The tool's own voice software plays these callers and calls the agent. Each conversation is recorded and then scored, usually by another language model that follows a checklist, plus fixed checks, such as how long the agent took to reply.
Say a health insurer sets up a voice agent that calls customers whose policies are about to lapse. The team writes 40 scenarios with five personas each: 200 test calls. The agent handles 184 of them correctly. In 9 calls, it quotes last year's premium of ₹14,200 instead of this year's ₹15,600. In 5 calls, it keeps talking when the caller interrupts. In 2 calls, it promises a discount that it is not allowed to offer. The team corrects the agent's instructions and its access to the premium table, and runs all 200 calls again. Every later change to the agent's instructions or its model is tested the same way before it reaches customers.
Voice agent testing sits between two neighbours. LLM evaluation scores a model's answers to written questions. Voice testing adds speech, timing and interruptions, which matter on a phone call. After release, automated quality assurance scores every real call. Voice testing scores imaginary calls, before release. The agents themselves are voice bots, often built on real-time voice AI that replies within a fraction of a second. For phone menus with fixed scripts, contact centre testing is enough.