TII introduces Falcon-ASR, an Arabic speech model focused on Emirati dialect

The Technology Innovation Institute (TII) in Abu Dhabi has introduced Falcon-ASR, a 1.6-billion-parameter speech recognition model focused on Arabic and the Emirati dialect. TII says the model also supports English, French, Spanish and Portuguese. Its training included Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English. The team says its aim is to transcribe everyday speech, including dialectal forms and changes between languages. Falcon-ASR also provides word-level timestamps, linking each transcribed word to its position in the audio.

TII describes Arabic speech as varying with region, speaker and recording setting. It notes that a system able to transcribe a formal broadcast may still have difficulty with Emirati conversation or phone recordings. The post says dialectal Arabic has fewer transcribed resources than Modern Standard Arabic, complicating both training and evaluation. Training included background noise, overlapping speech, music, room reverberation, telephony effects, and changes in speed and pitch.

Across six Arabic test sets, Falcon-ASR recorded an average word error rate (WER) of 20.92%. TII compared that result with a 23.17% best published score in the leaderboard snapshot it used. The team said Falcon-ASR’s average WER was 2.25 percentage points better than that result. The leaderboard calculates an equal-weight average across its six test sets, with lower error rates indicating better performance. In TII’s internal Emirati evaluation, Falcon-ASR recorded 22.73% WER and 10.19% character error rate (CER). The evaluation used held-out recordings and transcripts checked by people. TII says the model had the lowest WER and CER among the systems compared, with its WER 4.07 percentage points below Qwen3-Omni, the next-best result.

For English, TII reports a mean WER of 5.74% across seven public test sets used by the Hugging Face Open ASR Leaderboard. The five languages use the same model weights, and TII says users do not need to specify a language flag. For builders assessing multilingual transcription, the reported Arabic and Emirati evaluations offer performance measures to consider alongside the model’s language coverage. Teams can try Falcon-ASR with their own recordings in the Hugging Face Demo Space.