What happened
On 7 October 2026, the Technology Innovation Institute (TII) in Abu Dhabi published “Introducing Falcon ASR” on the Hugging Face blog. It describes Falcon-ASR as a 1.6 billion-parameter speech recognition model for Arabic, with a focus on Emirati. The same weights also transcribe English, French, Spanish and Portuguese, and the post says no language flag is required.
Key details
All results below are TII’s own.
- Arabic: an average word error rate (WER) of 20.92% and character error rate of 8.79% across six test sets of the Open Universal Arabic ASR Leaderboard. TII says the best published competitor it compared, Audar-ASR-V1-Turbo, scored 23.17% WER, so Falcon-ASR is “2.25 percentage points better”.
- Emirati: 22.73% WER and 10.19% CER on an internal TII evaluation, with Qwen3-Omni-30B-A3B-Instruct next at 26.80% WER. Because this evaluation is internal, it cannot be checked externally.
- English: a mean WER of 5.74% across seven Open ASR Leaderboard test sets, from 1.75% on LibriSpeech clean to 11.86% on Earnings-22.
- Not in the post: a license and inference speed or latency figures.
Why it matters
Dialect coverage is a known weak point for speech recognition. A model that targets Emirati Arabic and reports leaderboard-based comparisons for standard Arabic gives developers in the region a new option to test.
The strongest claim, about Emirati, rests on an internal evaluation, and the missing license means commercial use is unclear until the model card is checked.
Disclosure: Claude, made by Anthropic, was one of the AI tools used to research and draft this article. Claims and results are TII’s own.
Sources
- Introducing Falcon ASR Primary source , Technology Innovation Institute (Hugging Face blog)