Skip to main content

Voice Simulation

Why Voice Simulations Matter​

Voice simulations are synthetic, controllable reproductions of real customer calls. They have become a critical capability for any organization deploying AI voice agents. Three converging trends explain why.

1. Voice agents are shipping faster than teams can manually QA them​

Audio-native models have moved from 30% task-completion to 67% in eight months. That pace means new model versions, prompt changes, and integration updates ship constantly. Manual testing cannot keep up. Voice simulation provides repeatable, automated coverage that runs on every change.

2. Text-passing does not guarantee voice-passing​

An agent that solves a task flawlessly over text may fail the same task over voice. The gap comes from:

  • ASR errors compounding across turns. Accents, background noise, and telephony compression introduce transcription drift that accumulates as the conversation progresses.
  • Turn-taking failures. Interruptions, backchannels, and overlapping speech confuse tool-calling sequences and break the agent's conversation logic.
  • Integration fragility. A misheard account number causes an API call to fail, and the agent never recovers.

These are integrated failures that only surface when task execution and conversational dynamics are tested together, under realistic audio conditions. Benchmarking text and audio separately misses them entirely.

3. Customer experience assurance demands proactive, not reactive, quality​

Enterprises can no longer afford to let customers be the ones who "test" their systems. Voice simulation enables organizations to identify failures before they reach production, across AI agents, backend integrations, and handoffs to human agents.

What Okareo Does​

Okareo orchestrates real conversations against your voice agent over phone, WebRTC, SIP, or any voice channel, and evaluates what happened. The underlying transport is a configuration detail. Your testing practice is the same regardless of channel.

Voice Simulation runs the same engine as Multi-Turn Simulation, extended for live voice channels. If you already run text-based simulations, voice works the same way. Point at a voice target instead of a chat endpoint.

How a Simulation Works​

Every voice simulation follows a five-step lifecycle:

  1. Compose the building blocks. You define a Target (your voice agent), a Driver (a simulated caller persona), a Scenario (one or more test rows with inputs and expected outcomes), and Checks (the metrics that score the conversation).

  2. Start a Simulation. Okareo pairs one Target, one Driver, and one Scenario row. It initiates the call, applies the scenario parameters to the driver prompt, and begins the conversation with the configured first speaker.

  3. Real-time conversation. Once the call connects, both parties communicate simultaneously over a full-duplex audio channel. The Driver and Target can speak, interrupt, overlap, and respond naturally, just like a real phone call. Augmentations (background noise, barge-in, backchannel, secondary speakers) fire concurrently according to their configured probabilities. Stop conditions are monitored continuously and partial check results accumulate throughout the conversation.

  4. Stop and finalize. The run ends when the max turn count is reached, the Driver concludes, or a designated stop check returns true. Checks are computed a final time and aggregated across the conversation.

  5. Inspect. You get the full transcript, per-turn check annotations, final scores, the merged WAV recording, and (for voice targets) the raw audio per turn. Everything is available in the Okareo app and via the SDK.

Full-duplex voice simulation: Driver and Target speak simultaneously over a bidirectional audio channel while augmentations fire concurrently and stop conditions are monitored continuously, with results flowing into the Inspect phase (transcript, recording, checks)Full-duplex voice simulation: Driver and Target speak simultaneously over a bidirectional audio channel while augmentations fire concurrently and stop conditions are monitored continuously, with results flowing into the Inspect phase (transcript, recording, checks)

Use Cases​

A. AI Agent Validation​

Does your agent complete tasks, stay on policy, ground answers in your knowledge base, and resist jailbreaks? Simulate diverse caller personas across accents, speaking styles, and emotional states. Measure task completion, compliance, accuracy, guardrails, and bias.

AI Agent & Integration Testing →

B. Regression Testing Under Realistic Conditions​

Run the same test suite before and after a model change, under background noise, barge-in, and accent diversity. Controllable audio augmentation is how you catch regressions that clean-audio tests miss.

Voice Augmentation → | Experimentation and A/B Testing →

C. Load Testing​

What breaks when 50 calls land simultaneously? Latency cliffs, queue saturation, and routing failures. Load testing stress-tests the entire stack: network, routing engine, AI agent, CRM integrations, and agent desktop.

Load Testing →

D. Network and Voice Quality Monitoring​

Simulated calls placed under specific conditions detect carrier-level degradation, codec issues, and routing failures that only manifest on certain paths.

Voice Monitoring →

E. AI-to-Human Handoff Testing​

The highest-stakes decision a voice agent makes. Does escalation fire when it should, and only then? Premature escalation wastes agent minutes; missed escalation burns customer trust.

AI-to-Human Handoff Testing →

F. Continuous Production Monitoring​

Beyond one-time test passes, scheduled simulations catch drift, carrier outages, and integration regressions in production over time.

Monitoring → | Scheduling →

What's in This Section​

Everything below is available through both the Okareo app and the SDK. Start in whichever fits your workflow.

PageWhat you learn
Your First Voice SimulationRun a live scored call in two minutes
Personas and ScenariosCustom callers with voice identity and tone, multi-row scenarios, recordings
Voice ChecksCode, LLM, and audio checks; latency percentiles
Re-Scoring Past RunsApply new checks to an existing run without placing new calls
Voice AugmentationBackground noise, barge-in, backchannel, and other realistic conditions
Experimentation and A/B TestingStatistical comparison of two agent configurations with Bayesian and frequentist testing
CI GatingThreshold-based quality gates for CI pipelines
Load TestingConcurrent calls, p50/p90 latency under volume
AI Agent & Integration TestingWhat to test: task completion, compliance, accuracy, guardrails, bias, backend integration
AI-to-Human Handoff TestingValidating escalation decisions in both directions
Runnable scripts

Each page links to a companion script from the voice simulation cookbook where applicable.

Prerequisites​

  • An Okareo account (sign up)
  • A voice target: a phone number for your agent, a WebRTC endpoint, or a SIP URI

For SDK usage, also set OKAREO_API_KEY as an environment variable and install the Python package (pip install okareo).

Okareo handles the call orchestration, simulated callers, recordings, and quality scoring.

Cross-References​

  • Multi-Turn Simulation: shared concepts (Target, Driver, Scenario, Check) and chat / prompt / custom-endpoint patterns.
  • Creating Drivers: driver prompting theory (helper bias, drift, hard rules, turn-end checklist).
  • Scheduling Simulations: run voice simulations on cron, in CI, or on push.
  • Voice Monitoring: production observability for the same voice agents you tested here.
Cookbook

Full runnable scripts: tutorials/voice-simulation/