Why Callingtap · Quality
Voice quality, stated honestly
What speaks today, what is still being proven, and what we deliberately do not claim. No uptime percentages, no latency badges — only the stack and its current state.
What speaks today
The live voice path is the self-hosted stack: LiveKit for real-time audio, Whisper for speech recognition, and Kokoro for speech synthesis, running on Callingtap's own infrastructure at lk.callingtap.com. English and Hindi agents are live on this stack today.
Arabic: benchmarked, not yet routed
Arabic voices were benchmarked on the production server with Piper, which synthesised faster than real time. Routing is still off while the voice quality verdict is pending, so Arabic is honestly described as on the way — not live. Kokoro, the current live engine, has no Arabic support.
Turn-taking and interruptions
The pipeline includes turn detection, so callers can interrupt and the agent yields and resumes instead of talking over them. This is part of the shipped voice pipeline, not a demo effect.
What we do not claim
You will not find an uptime percentage, a latency badge, or a “sounds human” score on this page, because none of those has been measured and published for this stack. When a figure is measured, it will be shown with its measurement — not before.
Hear it yourself
The demo agents page lets you hear recorded sample conversations and, where live mode is available, talk to a real agent. Sample and live are labelled separately, never mixed.
Talk to the demo agents