Every score, client report and AI agent in Intella Studio starts with a transcript. Here is how accurate ours is on Arabic dialects, how we measured it, and against which engines.
Word accuracy in real-world Arabic environments — a 4.27% Word Error Rate, measured across 25 dialects, not curated studio speech.
Head to head
Arabic speech accuracy, measured on the same test set.
Every engine was scored on the identical 30,000-hour, 25-dialect Arabic corpus. Higher is better — this is word accuracy (100% minus Word Error Rate).
+29.5 points over the next-best global engine on real Arabic dialects.
How we test
A methodology built for the real world, not the lab.
Accuracy claims only mean something when the test conditions are honest. Ours are stated in full.
30,000 hours
Publicly available Arabic voice audio spanning 25 distinct dialects — Gulf, Egyptian, Levantine and more. Accuracy is measured on a part of it that was kept out of training.
40–110 dB SNR
Tested from quiet rooms to noisy contact-centre and field conditions, not just clean studio recordings.
20 Hz – 20 kHz
Full audible spectrum, covering the range of real telephony and microphone capture.
Word Error Rate
The industry-standard WER formula — accuracy reported as 100% minus WER, on the same corpus for every engine.
For BPOs
What a wrong transcript costs a BPO
When the transcript is wrong, the quality score is wrong, the client report is wrong, and the AI agent answers the wrong question.
Quality
A score your agents can accept
A check marked on the wrong words is a dispute waiting to happen.
Clients
Numbers you can defend
A client who listens to the call should hear what the report says.
AI Agents
Actions on what was really said
An AI agent can only act safely on what it understood.
Test Intella against your hardest calls.
Bring your noisiest, most dialect-heavy recordings. We'll show you the accuracy difference on data that matters to you.