Voho wins a landmark enterprise contract

Live demo

All demos

It starts speaking before it has finished reading.

The pause after a question is what makes an agent feel like a machine. This removes it.

Live microphone
ar-SA · streaming

What the caller said

Press the button to ask.

What it said back

Audio starts before the sentence is finished.

Retrieved from your documents
waiting

Returns & Warranty Policy (AR)

p. 4, §2.1 · match 0.91

Spare Parts Handbook 2026

p. 17 · match 0.84

Branch Operations Manual

p. 63 · match 0.62

Tool called
while speaking

The policy came from your documents. The dates did not — those are looked up live, mid-sentence.

lookup_purchase(invoice="INV-77120")
  → bought      2026-08-02
  → window ends 2026-09-01
  → 11 days left

This runs right here in your browser. Nothing to sign up for, and no call needed to see what it does.

The caller asks something that is answered in your own documents. Retrieval runs on what they have said so far, the model starts writing, and the first audio is on its way while the rest of the sentence is still being produced.

Build it yourself

Working example code, in Node and Python.

Clone yar-malik/realtime-arabic-voice-agent-najdi, paste in an API key, and the same thing you just played with runs on your machine.