TII releases Falcon-ASR, a 1.6B Arabic speech model with Emirati focus

The Technology Innovation Institute (TII) in Abu Dhabi has introduced Falcon-ASR, a 1.6 billion parameter speech recognition model for Arabic with a particular focus on the Emirati dialect. It also supports English, French, Spanish and Portuguese. All five languages run on the same weights, no language flag is needed, and the output is a transcript in the language spoken. The model also produces word-level timestamps that link each transcribed word to its position in the audio.

TII's pitch is that Arabic speech varies by region, speaker and setting. A model that copes with a formal news broadcast may still struggle with an Emirati conversation or a phone recording, and dialectal Arabic has fewer transcribed resources than Modern Standard Arabic (MSA), which makes training and evaluation harder. Falcon-ASR was trained on Emirati, MSA, other Gulf and Arabic dialects, and English, with the aim of transcribing everyday speech, including dialectal forms and switches between languages. Training data was augmented with background noise, overlapping speech, music, room reverberation and telephony effects, plus variations in speed and pitch, and the same treatment was applied to Emirati recordings. TII says this prepares the model for meetings, calls and other everyday recordings.

On Arabic benchmarks, TII followed the protocol of the Open Universal Arabic ASR Leaderboard, maintained by the ELM Research Center. That leaderboard ranks systems by the equal-weight average word error rate (WER) across six test sets and also reports character error rate (CER); lower is better for both. TII evaluated Falcon-ASR on the same six benchmarks using the leaderboard's pinned manifests and reports an average WER of 20.92%, versus 23.17% for the best published result in the leaderboard snapshot it used. That is a gap of 2.25 percentage points. Competitor figures are the published leaderboard averages as checked on 30 September 2026, not numbers TII re-ran.

Because public data already includes Emirati (the Casablanca dataset has a UAE subset), TII added an internal Emirati evaluation of extra Emirati and Gulf speech, using held-out recordings and human-validated transcripts. There Falcon-ASR scored 22.73% WER and 10.19% CER, the lowest WER and CER among the systems compared. Its WER is 4.07 percentage points below Qwen3-Omni, the next best result.

For English, on the seven public test sets used by the Hugging Face Open ASR Leaderboard, TII reports a mean WER of 5.74%. Falcon-ASR builds on TII's earlier Falcon3-Audio work, whose architecture and training approach are described in the paper "Competitive Audio-Language Models with Data-Efficient Single-Stage Training on Public Data".

A Hugging Face Demo Space lets people try the model with their own recordings. API access and native applications are planned.

Key facts

Falcon-ASR is a 1.6 billion parameter speech recognition model from TII in Abu Dhabi, aimed at Arabic with a focus on the Emirati dialect; it also handles English, French, Spanish and Portuguese with one set of weights and no language flag.
TII reports a 20.92% average WER across the six Arabic test sets of the Open Universal Arabic ASR Leaderboard protocol, versus 23.17% for the best published result in its snapshot, a gap of 2.25 percentage points.
On TII's internal Emirati evaluation, Falcon-ASR scored 22.73% WER and 10.19% CER, with WER 4.07 percentage points below Qwen3-Omni, the next best system compared.
On the seven public English test sets of the Hugging Face Open ASR Leaderboard, TII reports a mean WER of 5.74%.
A Hugging Face Demo Space is available now; API access and native applications are planned.

Why it matters

Dialectal Arabic has fewer transcribed resources than Modern Standard Arabic, and TII says that makes both training and evaluation harder. Falcon-ASR targets that gap, with particular attention to Emirati speech, and TII reports a lower average WER than the best published result in the leaderboard snapshot it used (20.92% against 23.17%). The model builds on TII's earlier Falcon3-Audio work.

Who it affects

The model is aimed at people who need transcripts of spoken Arabic, especially Emirati and other Gulf dialects, including speech recorded in meetings, calls and over phone lines. It also covers English, French, Spanish and Portuguese, so mixed-language recordings are within its stated scope.

How to use it

TII offers a Hugging Face Demo Space where you can try Falcon-ASR on your own recordings. The model returns word-level timestamps, and the same weights handle all five languages without a language flag, returning a transcript in the language spoken. API access and native applications are planned; the post gives no timescale. No licence, weights download link or open-source status is stated; only the Demo Space is mentioned.

How solid is it

This is a first-party blog post, and the results are TII's own evaluation. The Arabic evaluation follows the Open Universal Arabic ASR Leaderboard protocol with its pinned manifests, but the competitor figures are published leaderboard averages checked on 30 September 2026, not re-run by TII. The internal Emirati evaluation uses held-out recordings and human-validated transcripts, but it is not public, so others cannot reproduce it. No independent verification is mentioned.

Risks and caveats

The headline 2.25 percentage point gap is measured against a single leaderboard snapshot, and the post does not name the other systems on the Arabic leaderboard. Qwen3-Omni's own score on the Emirati evaluation is given only as a 4.07 percentage point gap. No per-test-set Arabic scores are given, and the English 5.74% is not compared with any other system. No scores are reported for French, Spanish or Portuguese. TII also gives no training data size or compute figures.

“Our aim is to transcribe the words people use in everyday speech, including dialectal forms and changes between languages.”

— TII, Falcon-ASR announcement