Zero-latency Arabic voice and document OCR directly into banking apps

A Saudi retail bank built Arabic voice and document capture straight into its mobile app with the Sarj SDK. Customers now open accounts and complete service requests by speaking, while the app fills in forms and reads their documents in real time. Voice responses stay under 120 milliseconds, and abandoned onboarding sessions fell by 40%.

The bank's app was already its busiest branch. But the journeys that mattered most, opening an account or updating personal details, were long forms on a small screen, and many customers gave up halfway or called the contact center to finish.

The challenge

Digital onboarding asked customers to type a lot: personal details, employment, address, and the information printed on their ID. Each field was a chance to make a mistake or to leave.

Customers who got stuck called the contact center, which turned a self-service journey into an agent's time. The bank wanted the app itself to help, in Arabic, at the moment a customer hesitated.

Two constraints narrowed the options. A voice assistant inside the app had to respond as fast as a person, or customers would not use it. And as a bank, it could not send customer audio or document images to infrastructure outside the Kingdom.

Why Sarj

The bank needed one provider for voice and documents that it could embed rather than redirect to.

• An SDK, not a separate channel. Sarj's SDK runs inside the bank's own app and design system, so the assistant feels like part of the product.
• Speed. Recognition, reasoning and speech synthesis run on one streaming pipeline in the Kingdom, keeping responses under 120 milliseconds.
• Documents in the same flow. Sarj's OCR reads Arabic and English identity documents in the same session as the conversation.
• Hosting and compliance. Customer data and models stay in Saudi Arabia, and the deployment is aligned with SAMA and NCA requirements.

The solution

1. Voice inside the onboarding journey

Customers can tap the microphone at any step and answer in their own words. The assistant asks only for what is missing and fills in the form on screen as they speak, so they can see and correct each answer.

2. Document capture and OCR

When the journey reaches identity, the customer photographs their ID. Sarj reads the document, fills in the matching fields, and asks the customer to confirm anything it is unsure of, instead of making them type it again.

3. Real-time screen sync

The voice assistant and the screen stay in step. When a customer asks where to find a setting or how to raise a request, the app moves to the right screen and highlights the next action while the assistant explains it.

Results

From pilot to the first full quarter in production:

• Under 120 milliseconds end-to-end voice latency inside the app.
• 40% fewer abandoned onboarding sessions on journeys where customers used voice.
• 3 weeks from SDK integration to the first pilot release.
• No customer audio or document images leave the Kingdom.

Conclusion

Voice made the bank's longest journeys feel short. Customers explain what they need in Arabic, the app does the typing, and the contact center hears from fewer people who got stuck.

The bank is now extending the assistant to card services and to service requests that today still start with a call.