TII introduces Falcon-ASR with a focus on Emirati speech

Abu Dhabi’s Technology Innovation Institute introduced Falcon-ASR, a 1.6-billion-parameter speech recognition model focused on Arabic, particularly the Emirati dialect, in a post dated October 7, 2026. TII says it also transcribes English, French, Spanish and Portuguese.

The team’s post directs users to a Hugging Face demo to try their own recordings. It describes API access and native applications as planned, rather than announcing their availability.

The main performance claim is an average word error rate of 20.92% across six Arabic test sets. TII compares that with the best published average of 23.17% in a leaderboard snapshot checked on September 30—a difference of 2.25 percentage points. Lower is better. The institute says it followed the Open Universal Arabic ASR Leaderboard’s protocol, which gives each test set equal weight. The comparison is therefore tied to that snapshot, not an unrestricted claim to outperform every competing system.

For Emirati speech, TII reports a separate internal evaluation using held-out Emirati and Gulf recordings with human-validated transcripts. Falcon-ASR recorded a 22.73% word error rate and a 10.19% character error rate. TII says both were the lowest among the systems it compared, with word error 4.07 percentage points below Qwen3-Omni, the next-best result. Those figures describe TII’s own test, distinct from the six-set public benchmark comparison.

The dialect focus addresses a problem TII highlights: transcribing formal Arabic broadcasts is not the same task as handling Emirati conversation or a phone recording. The team says dialectal Arabic has fewer transcribed resources than Modern Standard Arabic. Its training included Emirati, Modern Standard Arabic, other Arabic dialects and English, with noise, overlapping speech, reverberation and telephone effects among the recording conditions used.

All five supported languages use the same model weights without a language flag, according to TII. The model also supplies word-level timestamps, linking words to their positions in the recording. For English, the institute reports a mean word error rate of 5.74% across seven public test sets.