Falcon-ASR takes on Arabic speech at 1.6B parameters: 20.92% average word error across six test sets, with Emirati dialect in focus

Falcon-ASR is a 1.6-billion-parameter speech recognition model centered on Arabic, with particular focus on the Emirati dialect, and it also supports English, French, Spanish and Portuguese. Arabic speech varies by region, speaker and recording setting, so a model that handles a formal news broadcast may still struggle with an Emirati conversation or a phone recording; dialectal Arabic also has fewer transcribed resources than Modern Standard Arabic, making training and evaluation harder. On the six test sets and pinned manifests of the Open Universal Arabic ASR Leaderboard maintained by the ELM Research Center, Falcon-ASR averaged 20.92% word error rate, 2.25 percentage points below the best published result of 23.17% in the leaderboard snapshot checked on 30 September 2026.

On an internal Emirati evaluation the model reached 22.73% word error rate and 10.19% character error rate, the lowest of both metrics among the systems compared. Public evaluation data covers Emirati speech only through the UAE subset of Casablanca, so the team added an internal evaluation using held-out recordings and human-validated transcripts. On that internal Emirati evaluation Falcon-ASR's word error rate was 4.07 percentage points below Qwen3-Omni, the next best result; the same weights averaged 5.74% word error rate on the seven public English test sets used by the Hugging Face Open ASR Leaderboard.

Training data spans Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English, and includes background noise, overlapping speech, music, room reverberation, telephony effects, and variation in speed and pitch. These are the recording conditions of meetings, calls and everyday audio, and the same treatment was applied to Emirati recordings. All five languages share the same weights without requiring a language flag, and the output is a transcript in the spoken language; the model also provides word-level timestamps linking each transcribed word to its position in the audio.

The demonstrated scope is the average word error rate on six public Arabic test sets, the team's internal Emirati and Gulf speech evaluation, and the average word error rate on seven public English test sets; the model can be tried in a Hugging Face demo space, while API access and native applications are planned.

The lowest Emirati word and character error rates come from the team's own internal evaluation, whose recordings and human-validated transcripts are not part of the public test sets; the public leaderboard comparison is against published results checked on 30 September 2026.