TII Releases Falcon-ASR: 1.6B Arabic Speech Model Hits 20.92% WER
MarkTechPost Asif Razzaq
TII in Abu Dhabi released Falcon-ASR, a 1.6B Arabic speech model. It leads some Arabic tests, but right now you can only try the hosted demo.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
The Technology Innovation Institute in Abu Dhabi has put out Falcon-ASR, a 1.6B-parameter speech model built for Arabic and tuned with a clear bias toward the Emirati dialect. It also handles English, French, Spanish and Portuguese. One set of weights covers all five languages, and the model can return word-level timestamps, which makes it more useful for subtitles, searchable recordings and meeting notes.
On TII’s own Arabic evaluation, Falcon-ASR posts a 20.92% average word error rate across six test sets. That puts it 2.25 points ahead of the best published entry in the leaderboard snapshot TII used, Audar-ASR-V1-Turbo at 23.17%. TII also says the model came out ahead on its internal Emirati and Gulf test set, where it scored 22.73% WER and beat Qwen3-Omni-30B-A3B-Instruct by 4.07 points.
The interesting part is that this isn’t just another Modern Standard Arabic system with a dialect sticker slapped on top. TII says it trained Falcon-ASR on Emirati, MSA, other Gulf and Arabic dialects, plus English, and used audio augmentation to make the model tougher: background noise, overlapping speech, music, room reverberation, telephony effects, speed changes and pitch shifts all went in. The goal is everyday speech, including code-switching.
English looks solid too. Across seven English test sets from the Hugging Face Open ASR Leaderboard, TII reports a 5.74% mean WER. The weakest English result was 11.86% on Earnings-22, while LibriSpeech clean came in at 1.75%. TII also says French, Spanish and Portuguese are supported, although it did not publish scores for those languages.
And then comes the catch. Falcon-ASR is only available as a hosted Hugging Face demo for now, with clips limited to 60 seconds and 20 MB. TII says API access and native apps are planned, but there are no published weights, no quantized variants and no disclosed license yet. So the model is impressive on paper, but developers who want to ship with it will have to wait a bit longer than the leaderboard crowd.
My take — AI-written commentary, not fact-checked reporting
This is the usual AI-age split screen: strong numbers for the blog post, very little for the people who actually need to build. Open weights are still the difference between a research win and a product people can ship. A demo is cute; a downloadable model is infrastructure.
Read more about this at: MarkTechPost