Speech recognition and language models have improved a lot in the last two years. Word error rates are lower, and models make fewer factual mistakes on measured tasks. The failures that reach callers now come more and more from the system around the model: the media connection, turn detection, tool calls, application state and recovery after timeouts. That changes what voice agent evaluation has to cover.

I used this illustrative example in my Voice and Conversations on the Net (VON) Fall ’26 conference talk on Evaluating Real-Time Voice Agents Beyond AI Models: a caller is updating their billing address with a voice agent. They say “14 Pine Street,” then correct themselves: “Sorry, make that 40 Pine Street.” The agent answers, “40 Pine Street, got it.” But the tool call reads an earlier draft of the address and sends 14 Pine Street to the account system.

This is why, when evaluating voice agents at WebRTC.ventures, we do not rely solely on the AI model or benchmarks. We look at the full production voice pipeline, including SIP/WebRTC audio paths, call handling, permissions, state, and error recovery.  In this post I walk through where those voice agent failures happen in a real-time call, how chasing lower voice agent latency can cut callers off, what evidence catches each failure, how to build a voice agent test set, and how to turn the results into a release decision.

You can also view my full VON 2026 conference presentation on this topic. 

The voice agent call path: 7 places it can fail

A voice agent is a chain of systems. Some architectures use separate speech-to-text, language and text-to-speech models. Others use a speech-to-speech model. Either way, every call passes through the stages below, and each one can fail while the model behaves correctly.

  1. Connection (SIP, RTP, WebRTC). Packet loss, jitter and buffer delay degrade the audio before any model hears it. Concealment fills gaps with synthesized audio, so speech recognition receives sound the caller never made.
  2. Turn detection (VAD and end-of-turn). An aggressive endpoint decides the caller has finished during a pause. The agent responds to half a request.
  3. Speech-to-text. Streaming transcripts change as more audio arrives. An agent that acts on a partial transcript acts on words the caller has not finished saying.
  4. Agent model and conversation flow. This is where most evaluation effort goes today, and where model progress has been strongest.
  5. Control layer: policy and tools. The application decides what the agent may do and which values it sends. The Pine Street failure happens here: the tool reads a stale draft.
  6. Backend. A tool call can time out after the update commits. The agent has no confirmation, and the account has already changed.
  7. Text-to-speech and return audio. Generated speech travels back over the same media connection. Late speech, or speech that keeps playing after the caller interrupts, breaks turn-taking.

These stages interact. When a caller interrupts, stopping the generated speech does not cancel a database update already in flight. When an API times out, the action may already have succeeded. The application needs rules for those situations, and voice agent evaluation checks whether the rules hold.

The model can also improve while the system regresses. A new model version, a different endpointing setting or a changed tool schema can each break a flow that passed last week.

Real-time voice agent architecture: SIP and WebRTC connection, turn detection, speech-to-text, agent, tools, text-to-speech and backend, with failure points at each stage.

Voice agent evaluation metrics across the full system

We organize voice agent evaluation around four dimensions: Conversation, Comprehension, Conduct and Completion. Each measures what the caller experiences, and each can break for reasons outside the model. Scoring them separately shows you what to fix. A single blended score hides different problems behind one number.

DimensionQuestion it answersWhat can break itEvidence and metrics
ConversationDid the conversation flow naturally?Packet loss, jitter, endpointing, interruption handling, late speechReceived audio and a speech timeline. Jitter, RTT, bitrate, packet loss, concealment, buffer delay, audio level
ComprehensionDid the agent understand the caller correctly?Audio quality, speech-to-text errors, partial transcripts, corrections lost in application stateCaller audio, transcript and retained details. WER, CER, entities, intent
ConductDid it take the right actions and follow the rules?Stale state, wrong tool arguments, skipped verification, actions in the wrong orderVerification events, tool inputs and results. Required steps, policies, tool contracts, action order
CompletionWas the caller’s goal achieved?Tool timeouts, backend errors, handoffs offered but never connectedVerified application state and required explanations. Goal, final state, claim vs. result, handoff, coverage

A low word error rate can hide the error that matters. In the Pine Street case, one substituted word in a 20-word request is a 5% WER, but that one word is an essential part of patient’s address.

Define success for each use case before you measure. For customer support, success might be a verified account change or a connected human handoff. For sales, accurate qualification and a correctly saved appointment. For a training agent, staying in the scenario and giving feedback against an expert-approved rubric, which a short call does not prove.

Four dimensions of voice agent evaluation: conversation, comprehension, conduct and completion.

Voice agent latency: why faster responses can cut callers off

Latency matters, but first check where the measurement starts. Time from the system’s end-of-turn decision to response audio leaves out how long the system waited before deciding the caller had finished. Shorten that wait and the number improves. The agent also starts answering callers who pause mid-request: “Cancel my subscription… at the end of the billing period.”

Measure from the caller’s actual speech boundary, and separate audio arriving from audio being played. Track p95 response time alongside premature replies and failed interruptions, because turn detection and interruption handling change together. For a training agent, giving the user time to think may matter more than shaving every pause.

Here is what that looks like in a pre-release comparison: the same billing-address test cases and checks, two agent versions, 100 test calls each.

DimensionFailure checkedCurrentCandidateLimit
ConversationCuts off an unfinished request2 / 10012 / 100≤ 3 / 100
ComprehensionReads back 14 after the correction to 400 / 1008 / 1000 / 100
ConductTool value differs from the confirmed address0 / 1004 / 1000 / 100
CompletionMissing required invoice explanation1 / 1006 / 100≤ 1 / 100

The candidate improves p95 response time from 1,100 ms to 780 ms, and both versions meet a 1,200 ms budget. The candidate still exceeds the failure limit in all four dimensions. These are synthetic numbers for illustration, not production results.

Apply the same discipline to cost. Compare cost per successfully completed task, alongside cost per minute. A cheaper configuration is worth it when it preserves the required outcomes. A configuration that saves inference cost by creating failed calls or avoidable handoffs costs more per completed task.

Voice agent tool calls: check what was saved against what was said

Hearing “40” in the recording establishes what the agent said. It does not establish what the agent saved. To catch the Pine Street failure, inspect the value submitted to the tool and read the updated account independently.

Use code to check recorded facts: which account changed, which address was submitted, and whether identity verification came before the write. Use an LLM judge for meaning: did the agent keep the caller’s correction, and did it explain the result accurately? Give the judge explicit criteria, then compare its verdicts with independent human reviews on calls you did not use to tune it. Anthropic’s evaluation guidance recommends combining both approaches. A schema-valid judge output can still be wrong, and no LLM assessment proves a backend record changed.

Completion needs two separate verdicts: did the task finish, and did the agent have evidence when it said so?

Claim supported by evidenceClaim made without evidence
Task completedVerified successUnsupported confirmation
Task not completedTruthful noncompletion and recoveryFalse success

False success is the result that hurts customers most: the agent tells the caller the plan is canceled, and it isn’t. Unsupported confirmation also counts as a failure, even when the backend happened to succeed, because the agent claimed something it could not know at that moment.

Two rules come out of this. A timeout does not tell you whether the action happened, so the agent has to reconcile the original operation before it confirms anything. Stopping speech does not undo a commit, so an interruption should suppress the stale response and then reconcile the effect. When the state is unknown, it stays unknown until the system checks.

How to build a voice agent test set

Start with one workflow where a failure is expensive, and define acceptable outcomes before you choose an evaluation tool.

  1. Choose a workflow with one costly failure, such as a billing address change.
  2. Define success: required behavior, forbidden behavior, and the evidence that proves each goal.
  3. Test the failures: the normal request, a correction before submission, an interruption, skipped identity verification, and a timeout after the update commits.
  4. Inspect and repeat: check the evidence, fix, and run the same cases again.

Use staging accounts and reset their state between runs. Run text-level tests often to check reasoning and tool use, then audio tests to exercise speech, timing and interruptions. LiveKit’s simulation documentation makes this distinction explicit. Before launch, test the actual browser or telephone path with representative languages, background noise, devices, network conditions and concurrent calls. Our QA team covers that layer with its QA framework for AI voice agents.

Open-source voice agent evaluation with VoiceAI QA

For my talk I built VoiceAI QA, an open-source workbench for this workflow (the repository is named Conversation Agent Evals). It works in three steps:

  • Run: a simulated caller built on Pipecat holds real voice conversations with your agent and saves the audio, transcript and available tool and state evidence.
  • Evaluate: deterministic code checks verify recorded facts, and an optional LLM judge (through the ASSERT framework) scores behavior against a rubric. Code findings and LLM assessments stay separate.
  • Compare: the same cases run against two agent versions, and the report shows regressions by dimension.

The important part is keeping each result linked to its test case, agent configuration and evidence. Re-evaluating an old recording with new criteria improves your grading. Testing a changed agent requires new conversations.

Some teams prefer a managed platform. Coval and Bluejay offer conversation simulation, automated evaluation and production monitoring, which reduces the infrastructure you maintain. We also help set up and integrate these platforms. Whichever tool you use, someone still has to define the scenarios, the success criteria, and the checks that verify what happened in your application and backend.

Release criteria for production voice agents

For an agent that changes accounts, some failures block a release regardless of average scores: a write to the wrong account, an action taken without verification, or a success claim without evidence. Product leaders should define acceptable outcomes and tradeoffs with engineering. Engineering should make those requirements observable and enforceable.

After launch, monitor completion, handoff success, latency, interruptions, tool errors and cost. Review failures by workflow and by calling conditions, then turn each one into a regression test. Sample real conversations too, because a passing simulation does not cover every production call.

Keep enough evidence to investigate, with appropriate access and retention controls. The IETF’s vCon draft defines a portable container for conversation data, including audio, transcripts, analysis and attachments such as tool traces. We covered how to capture these records in a healthcare voice AI agent built with LiveKit and vCon.

Voice AI integration and evaluation services

Catching a failure like the Pine Street address takes a team that works on the audio path, the agent and the backend together. WebRTC.ventures builds real-time voice AI across that pipeline, from WebRTC and SIP media to speech recognition, LLM orchestration and speech synthesis. Contact us to talk about your voice AI project.

Further Reading:

Recent Blog Posts