HumanStandard
Get early access to Vak
AI voice detection for live calls and recorded audio.
Vak returns a record of whether a voice is human or synthetic, in real time on a live call or after the fact on a recording. Access opens to a small group of teams first.
Vāk is Sanskrit for speech.
Live calls, in real time
Stream the caller’s channel and Vak returns a running verdict every half second. A confident answer arrives after two seconds of speech on clear audio.
Recorded calls and clips
Each side of a recorded call is scored from the audio alone, and a single clip returns a verdict on synthetic speech, including voices from engines Vak was never trained on.
Phone, web and compressed audio
Trained and tested on audio after telephone codecs, WebRTC compression and MP3 re-encoding, the conditions real calls arrive in.
Evidence
Measured results
- Synthetic voices flagged after 2 seconds of live speechStreaming test · clear audio · 1% false-positive rate
- 90.4%
- Same, after 4 seconds over a telephone lineStreaming test · G.711 phone codec · 1% false-positive rate
- 85.5%
- Compute per live updateOne GPU · verdict refreshed every 0.5 seconds
- 21 ms
- Synthetic voices flagged, full speaking turn25 text-to-speech engines withheld from training · 1% false-positive rate
- 93.0%
- Real voices flagged after MP3 re-encoding800 held-out human recordings · 64 kbps MP3
- 0.5%
- Agent side identified in public conversationsRecorded conversations between people and commercial voice assistants
- 56 / 66
Internal evaluation, 24 October 2026, on held-out test sets; live figures measured in streaming tests. Deployment performance is measured and reported for each customer setting.