Voice AI isn't judged by
what it says.
It's judged by how it sounds.
Benten is open-source observability for Voice AI. We listen to the actual conversation — the latency, the silence, the interruptions — so you hear what your customers hear.
MIT licensed · Self-hosted · No hosted service, no lock-in
- HumanHey, can you check my order status?
- AgentSure — one moment while I pull that up.
- HumanHey, can you check my order status?
- HumanHello?… Hello?
- SystemCall abandoned
Two identical sentences.
One feels human. One doesn't.
Two conversations can say the exact same words. The difference isn't language — it's rhythm. Benten measures the rhythm.
Instant
100msFeels like a person.
Conversational
400msYou barely notice it.
Slow
1200msSomething feels off.
Broken
2500ms"Hello?…"
Your transcript says everything went perfectly.
The customer remembers the silence.
Your customers never read the transcript. They hear the pauses, the interruptions, the dead air. Words are not the product. Timing is.
What the transcript shows
User: I'd like to change my flight.
Agent: Sure, I can help with that.
Looks perfect. Ship it.
What Benten hears
User: I'd like to change my flight.
Agent: Sure, I can help with that.
The customer already thinks the call disconnected.
Don't trust your provider.
Measure the conversation yourself.
Every voice provider grades their own homework. Benten is an independent layer of truth above ElevenLabs, Vapi, and Retell — so you can compare providers fairly, and see what actually shipped.
Your provider dashboard
Says: "Everything is healthy."
What Benten measured
Reality: Your customers are hanging up.
Your Voice AI lives or dies
before it finishes the sentence.
Every delay changes perception. 100ms feels instant. 400ms feels human. 1,200ms feels slow. 2,500ms and they think the line dropped.
Voice has no observability.
Until now.
Cloud providers monitor servers. Voice providers monitor APIs. Neither monitors conversations — and conversations are what your users actually experience.
Every pause tells a story.
Not because they're metrics.
Because they're the reason conversations feel human.
Your QA team never notices these — they're not crashes, they're half-second hesitations. The bugs your customers notice first are the ones your logs never record.
Dead air
The silence that makes callers think you dropped.
Interruptions
When the model cuts the human off mid-word.
Barge-ins
When the human has to talk over the assistant to be heard.
Latency
The gap from human-stop to agent-start, in milliseconds.
Noise
Line quality, distortion, and background artifacts.
Speech overlap
The two-people-talking-at-once moments.
Every provider tells a different story.
The conversation only happened once.
ElevenLabs says one thing. Vapi shows another. Retell exposes something else. Benten sits above all of them and gives you one timeline, one set of metrics, one truth.
ElevenLabs
Their version
Vendor-graded metrics
Vapi
Their version
Vendor-graded metrics
Retell
Their version
Vendor-graded metrics
Benten
One conversation. One timeline. One truth.
Every conversation leaves
a fingerprint.
Some feel effortless. Some feel awkward. Some make people hang up. Benten shows you exactly why — the shape of every call, the moment it went wrong, the millisecond it fell apart.
"That felt weird" has a cause.
Too much silence. Interruptions. Talking over people. Bad pacing.
Benten measures the moments that make conversations feel human.
No hosted service.
No lock-in.
Just the code.
Benten is 100% open source. Clone it, run it on your own infrastructure, and own your voice observability stack end-to-end. We don't run a cloud. We don't want your call audio.
$ git clone github.com/ayusrjn/benten $ cd benten $ docker compose up → syncing agents from ElevenLabs... → syncing agents from Vapi... → syncing agents from Retell... ✓ 34 agents discovered ✓ listening to conversations ✓ dashboard live at localhost:3000