Editorial illustration for TII's Falcon-ASR Arabic Model Shows 20.92% WER, Trails Qwen3-ASR
TII's Falcon-ASR Arabic Model Shows 20.92% WER, Trails...
The Technology Innovation Institute in Abu Dhabi has put out Falcon-ASR, a 1.6 billion parameter speech recognition model built around the Emirati Arabic dialect. The model also handles English, French, Spanish and Portuguese, all through one shared set of weights, with no language flag needed. On TII's own benchmarks, Falcon-ASR posts a 20.92% average word error rate across six Arabic test sets, which the institute says beats the best entry on its leaderboard snapshot, Audar-ASR-V1-Turbo, by 2.25 points.
That's the headline number. What's missing is just as notable. TII hasn't published weights, a license, or quantized versions, and there's no API yet, only a hosted demo on Hugging Face that caps clips at 60 seconds and 20 megabytes.
Every score in the comparison is self-reported by TII, with no independent verification so far. For developers weighing Arabic voice tools, that leaves a gap between what the benchmarks claim and what's actually available to build on right now.
In TII’s evaluation, Falcon-ASR posts a 20.92% average word error rate (WER) across 6 Arabic test sets. That beats the best published entry in TII’s leaderboard snapshot by 2.25 points.
Why this matters The headline number, a 2.25-point improvement over the "best published entry," sounds like progress until you notice Qwen3-ASR-1.7B is sitting at 5.76% mean WER on the same leaderboard, less than a third of Falcon-ASR's error rate. TII picked its comparison set carefully, and that's worth flagging for anyone deciding which model to build on. For developers targeting Emirati Arabic specifically, a dialect-focused 1.6B model beating generic leaderboard entries by two points is a real, if narrow, win.
But the restricted access TII describes limits how fast anyone outside Abu Dhabi can actually test that claim against production traffic. The silence on French, Spanish, and Portuguese scores is its own signal: TII is happy to publish numbers when they look good and stay quiet when they don't. Our read: this is a solid regional tool, not a general-purpose ASR leader, and teams evaluating Arabic speech models should run Qwen3-ASR and Falcon-ASR side by side on their own dialect mix before trusting either leaderboard snapshot.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives AI agent teams waste tokens for barely measurable quality gains, research finds
From Archives Sakana AI's agentic LLM peer review catches 73% of core-claim errors
Further Reading
- Papers with Code Benchmarks - Papers with Code
- Chatbot Arena Leaderboard - LMSYS
