AI delivery
Gulf dialect speech and annotation programme for a global AI developer
A large AI developer building multilingual foundation models, headquartered in North America.
- Industry
- Technology and Startups
- Pillar
- AI delivery
- Client geography
- North America
- Delivery location
- UAE remote workforce with Corpshore AI global quality layer
- Languages
- Arabic C2 across Gulf, Levantine, Egyptian and Maghrebi dialects, English C1
- Engagement model
- Programme-based, output priced
- Timeline
- Nine months across three phases, extended
The challenge
The client's models performed acceptably in Modern Standard Arabic and poorly in the dialects people actually speak. Gulf Arabic was the weakest of the major dialect groups in their evaluation set, and the training data to fix it did not exist in usable quantity or quality. Previous vendor attempts had produced data that failed acceptance, mostly because annotators from one Arabic-speaking region had been used to label speech from another, and dialect judgements had been made by people without native intuition for the variety in question.
What Corpshore did
We recruited and trained a distributed workforce of native dialect speakers, matched by dialect rather than by language. Emirati, Saudi, Kuwaiti, Qatari, Bahraini and Omani speakers worked on Gulf material. Levantine, Egyptian and Maghrebi material was routed to native speakers of those varieties.
The programme ran in three phases: field speech collection under controlled conditions with consented speakers, transcription and dialect annotation including code-switched Arabic and English speech, and model output evaluation for linguistic accuracy, register and cultural appropriateness.
Quality was managed through layered review with measured inter-annotator agreement, an adjudication tier for disagreements and a standing feedback loop into guideline refinement. Where the guidelines produced inconsistent outcomes, we said so and proposed revisions rather than delivering data we knew was unstable.
Delivery model
Programme-based with output pricing, remote UAE and regional workforce, Corpshore AI quality operations layer, full consent and data governance chain for all collected speech.
Results
- Delivered hundreds of hours of transcribed and annotated dialect speech across the first three phases, all passing the client's acceptance criteria.
- Inter-annotator agreement held above the programme threshold throughout.
- The client reported measurable improvement in dialect performance on their internal evaluation set following training on the delivered data.
- The programme was extended twice and expanded into additional Arabic varieties.
Why it worked
Dialect competence is not transferable within a language. Matching annotators to the specific variety, rather than treating Arabic as one skill, was the difference between data that passed acceptance and data that did not.
Discuss an engagement like this
All UAE enquiries answered within six hours.
This is a representative engagement. The client is anonymised, and the figures are drawn from engagement reporting. We do not name a client without written consent, and we do not publish a number we cannot evidence.