Voice agents fail in ways a transcript can’t show. A reply that lands two seconds late, a name said wrong, a caller cut off mid‑sentence. Netra now tests your agent against the callers you actually get, and listens to every call it takes.
Build a library of personas, each with its own voice, pace, clarity, accent and temper. Netra calls your agent as each of them, captures the conversation and evaluates it, so a release is tested against the people who will actually phone in.
How personas work (opens in new tab)Frustrated caller en-US
“This is the third time I’ve called. I was charged twice, and nobody called me back.”
What this caller tests
Does your agent stay calm, acknowledge the problem and keep its turns short?
Fast talker en-IN
“Can you move the delivery to Friday and send it to my office instead of home?”
What this caller tests
Does speech-to-text keep up when every word arrives at once?
Non-native speaker es-MX
“I want to… how you say… cancel the second order, please.”
What this caller tests
Does your agent understand accents, and stay in the caller’s language?
Rambler en-GB
“So, um, it started last week, no, the week before, when the, uh, app logged me out.”
What this caller tests
Can your agent steer a wandering caller back to the point without talking over them?
Confused caller en-AU
“Sorry, which number was that again? The one on the card?”
What this caller tests
Does your agent slow down, repeat clearly and say the numbers right?
On a noisy line en-US
“Hi, can you hear me? I’m calling about my refund.”
What this caller tests
Does the conversation still work with traffic and chatter behind the caller?
Voice evaluators score the audio, not just the words. They cover the three places a voice agent breaks: the speech it produces, the speech it hears, and the conversation in between.
STTcaller turn · en-IN
Word error rate, transcription correctness and time to first transcript, on every turn the caller speaks.
Conversationstereo call · 38 s
Goal fulfilment, answer quality, tool calls, toxicity and bias, judged on every turn.
TTSagent reply · 3.4 s
Humanness, pronunciation, speaking rate and language accuracy, scored on every reply your agent speaks.
Both sides of the call on separate tracks, a live transcript, and a span for every stage between the caller speaking and your agent answering. When a turn goes wrong, you see exactly where the time went.
Transcript real time
Hi, I was charged twice for order 4471.
I'm sorry about that. Let me pull up order 4471.
It's the second time this month.
I see. I can see both charges on the account.
So can you refund one of them?
Yes, I've refunded the duplicate charge.
How long will that take?
Three to five working days. Anything else?
Speech-to-text, the model, tools and speech back out as separate spans, with the audio, the transcript and the timings lined up turn by turn. Works with LiveKit, WebSockets and Vapi.
Speech-to-text, the model, tool calls and text-to-speech each get their own span and latency, so you see which stage was slow.
STT, LLM and TTS, separately: Docs (opens in new tab)Every production call captured end to end, from the caller's first word to your agent's last reply.
Trace every call: Docs (opens in new tab)Turn-by-turn audio: what was heard, what the agent decided and what it said, with timings you can step through.
Every turn in detail: Docs (opens in new tab)Stereo recording with caller and agent on their own tracks. Silences, overlaps and interruptions are measured from the audio, not the text.
Audio-based data: Docs (opens in new tab)Every turn transcribed as it happens and lined up against the audio it came from.
Real-time transcription: Docs (opens in new tab)Add Netra.init() to a LiveKit agent in Python or TypeScript, or stream any voice stack over WebSockets. Vapi agents connect from the dashboard for simulation.
Start today
Rehearse with your first personas today, and evaluate every call after launch. Free to start.
