Voice traces, not text traces with audio bolted on
Netra traces voice pipelines as a first-class shape: transcription, reasoning, tool calls and synthesis land in a single span tree for the turn.
By agent type · Voice
A voice agent is three systems. The failure is usually the seam.
Speech to text, the model, text to speech. Each leg logs its own success, and none of them owns the call the caller actually had. Netra captures voice traces as one turn, so the mis-transcription that became a wrong answer is visible as a chain rather than three green ticks.
One turn, end to end
Latency, leg by leg
Per-stage timing tells you which part is making the caller wait — the recogniser, the model, the tool call or the voice coming back.
Scored on the transcript
The evaluators you already run on text — faithfulness, correct escalation, on-policy tone, PII leakage — applied to what was said on the call.
Barge-in and turn-taking, in order
The trajectory keeps the sequence of the conversation, so interruptions and re-prompts read the way the caller experienced them.
Cost per call
Recognition, tokens and synthesis attributed to the call and the route that caused them, rather than to three separate bills.
Alerts that link to the call
A quality or latency threshold fires into Slack or PagerDuty with the session attached — no grepping for the recording.
The failures that only voice has
It heard something else
An account number, a drug name or a street misheard by one syllable produces a confident, fluent, entirely wrong turn. The transcript sits next to the answer, so the cause is one glance rather than one theory.
Dead air
A tool call that takes four seconds is a shrug in chat and an abandoned call on the phone. Per-stage latency puts a number on it.
Right answer, wrong delivery
Reading a twelve-item list aloud is a failure even when every item is correct. Grade the response for the channel it is going out on.
Rehearse the call before you take it
Simulation puts personas through your scenarios — the frustrated caller, the ambiguous request, the one who interrupts — and scores each run with the same evaluators that watch live traffic. Red teaming runs the adversarial passes before a customer finds them.
Start today
See a full voice turn end to end
STT, reasoning and TTS in one span tree. Start free.