Hang up on voicemail before the beep.
Streaming answering machine detection for outbound calls. Human, machine or undecided, in the first seconds of the call and always before your deadline.
Every greeting you listen to is billed.
A dialer that cannot tell a person from a voicemail pays for the greeting, and for whoever is waiting on the other end.
Contact centers
Agents say hello to greetings, and stop dialing while they do.
Voice AI
Speech-to-text, the LLM and text-to-speech run on every second of a greeting.
Every dialer
Carrier minutes are billed from the answer: greeting, beep and the silence after it.
At 20,000 answered calls a day, about $5,040 a month goes to listening to machines. Run your numbers
Hear it decide.
Four listeners hear the call at once. The engine weighs what they report every 250 ms and answers as soon as the evidence agrees.
It sounds like a person, but it never stops to listen.
Listening 0.00 s
Synthesized audio at 8 kHz through G.711 µ-law. Timings are illustrative and follow how the engine decides.
- Median time to verdict
- 2.4 sTBD
- From answer to verdict, across all verdicts.
- Voicemails cut before the beep
- 94%TBD
- Greetings ended before the record tone.
- Verdict accuracy
- 97%TBD
- On a held-out set of labeled calls.
- People treated as machines
- < 1%TBD
- Live persons sent down the voicemail path.
Values marked TBD are placeholders until the benchmark is published.
You set the deadline. It answers inside it.
Pick a mode per stream and send a deadline. The engine answers 500 ms before it: with a verdict, or with an explicit undecided your dialer can act on.
Mode
The default. Waits for a confirming window before it calls a person a person.
The hatched end of the budget is the 500 ms answer margin. Illustrative timings.
Fits the telephony you already run.
Stream from Asterisk, FreeSWITCH, gRPC or WebSocket in µ-law, A-law or linear PCM, at 8 or 16 kHz. Not streaming? Send a recording over REST and get the same verdict.
; Answered calls enter Stasis; vmd-connector snoops the channel over ARI and streams
; the callee audio to the engine (chan_websocket, AudioSocket or RTP external media).
[outbound-amd]
exten => s,1,Set(VMD_MODE=balanced)
same => n,Set(VMD_DEADLINE_MS=4000)
same => n,Stasis(vmd) ; returns with VMD_LABEL and VMD_REASON set
same => n,GotoIf($["${VMD_LABEL}" = "human"]?agents,s,1)
same => n,Hangup() ; machine or undecided: free the trunkAPI v1 preview. Read the Asterisk guide
Priced per call that gets a verdict.
1,000 calls free every month, then from $0.004TBD per call. No per-minute meter, no setup fee.
Asked by people who run dialers.
Does it work on 8 kHz telephone audio?
Narrowband is home turf: G.711 µ-law, A-law and linear PCM at 8 kHz, the way calls cross real trunks. Wideband 16 kHz streams are supported too.
Which languages does it support?
It listens to how the call behaves rather than to its words: voice activity, turn-taking, tones, synthesized speech and acoustic fingerprints. That is not tied to one language. Known-greeting libraries are tuned per region.
Do you keep call audio?
By default, audio is processed in memory to produce the verdict and is not kept. Keeping recordings for review is an opt-in workspace setting.
How do we evaluate it before switching?
Run it in shadow mode next to your current AMD: it logs every verdict and acts on none, so you can compare call by call. Or send last month’s recordings over REST for a backtest. Both use the free quota first.