TII released Falcon-ASR on 7 October 2026 on Hugging Face, one day after it announced a three-model Arabic launch from Abu Dhabi. The launch also included Falcon-Emirati, a 7-billion-parameter language model, and Falcon-OCR-Arabic, which reads Arabic text from images and documents. Falcon-ASR is the speech model: 1.6 billion parameters, built on TII's earlier Falcon3-Audio work, and designed to turn speech into text in Modern Standard Arabic, Emirati Arabic, English, French, Spanish and Portuguese.
Users do not have to set a language before they start. The model returns a transcript in whatever language is spoken, and it can attach a timestamp to each word. TII names subtitling, searchable recordings, meeting and interview transcription, and accessibility tools as the intended uses. It says it trained the model on Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English, with added background noise, overlapping speakers, music, room echo, telephone effects and changes in speed and pitch.
Where the numbers come from
TII's public Arabic test covers six sets and reports an average word error rate of 20.92%. It sets that against 23.17%, the best published result in the leaderboard snapshot it used, which it says it checked on 30 September 2026. The public Arabic data already includes a UAE subset from the Casablanca benchmark. TII then added an internal test of Emirati and Gulf speech, using held-out recordings and human-checked transcripts. On that test, Falcon-ASR scores 22.73% word error rate and 10.19% character error rate.
TII's table puts Qwen3-Omni next, 4.07 percentage points behind on word error rate. TII's release also says Falcon-ASR surpassed a 30-billion-parameter multimodal model on Emirati speech. Both comparisons come from TII's own internal test, and TII chose which systems to include. On seven public English test sets, the Hugging Face Open ASR Leaderboard set, TII reports a mean word error rate of 5.74%.
RuntimeWire reports that TII's team considers Emirati Arabic underrepresented in transcription data, which is the reason given for the focus. Falcon-Emirati, the language model launched alongside the speech model, scores 84.83% on a benchmark of Emirati dialect and cultural knowledge, according to RuntimeWire. That is a language benchmark, not a speech result, and it does not measure transcription at all.
What a team can do now
The Hugging Face demo space is the only way to try the model today. TII says API access and native apps are planned but not yet available, and the sources checked give no price and no licence. Coverage from RuntimeWire notes that transcription quality can shift with vocabulary, speaker and recording conditions, and the internal Emirati result is the only one that targets everyday Emirati speech. Teams that need Arabic transcription for UAE customers should test the demo on their own audio before planning around it.