Khulasa

Khulasa is a meeting bot for bilingual teams. It joins a scheduled Google Meet, Zoom or Teams call, records it, and returns a transcript plus a summary that names the decisions, the risks and the action items — in Arabic and English, including the switches between them mid-sentence.

The business case

In a bilingual organisation the written record drifts from what was actually said. Minutes get written by whoever remembers, English terms come back mangled into Arabic script, and three weeks later nobody can tell who owned a decision. Khulasa closes that gap: every call ends with a short artefact — what was decided, who owns it, what is at risk — instead of an hour of audio nobody will replay.

Access is controlled at two levels: organisation-wide roles, plus permissions on individual recordings. A board call and a standup do not have to live in the same bucket.

For regulated industries the usual blocker with meeting-notetaker SaaS is that the audio leaves the country. Khulasa runs on-premise or in your private cloud, so recordings stay inside your own estate — the same product, on your hardware.

Pricing is per seat with an hours cap, so cost stays predictable as you roll it out: Free (3 seats, 3 h/user/month), Team ($20/seat/month, 30 h), Business ($34/seat/month, 120 h), and custom pricing for private cloud and on-premise.

How it works

Speech recognition runs on Whisper large-v3 — OpenAI's model, MIT licence, open weights. The interesting part is not the model choice but the configuration: Khulasa runs it at a tuned reference setup rather than library defaults, and most of the accuracy gap on code-switched audio comes from exactly that.

The transcript keeps English terms in Latin script instead of transliterating them into Arabic letters. dashboard stays dashboard. That sounds cosmetic, but it is what keeps transcripts searchable and keeps keyword and glossary matching working downstream.

Measured on UN ESCWA meeting audio (80 utterances, August 2026):

Overall word error rate: 35.14

English words inside Arabic speech: +16.3 … +34.7 points vs competing systems

Arabic–English switch points preserved: more than 3×

Deployment comes in three shapes — hosted tiers, private cloud, or on-premise on your own hardware. Because the speech model has open weights, the on-premise build is genuinely self-contained: no per-minute call out to a third-party API, and no vendor to renegotiate with when volume grows.

Recordings and transcripts are encrypted in transit (TLS) and at rest, and every organisation's data is strictly isolated from every other. Meeting content is processed to produce your analysis and to run and secure the service — not to train models. Training is opt-in, and off by default.