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

The work is motivated by the wide variation in Arabic speech: a model that handles a formal news broadcast may still struggle with a conversation in Emirati or with phone-line audio. Dialectal Arabic also has fewer transcribed resources than Modern Standard Arabic (MSA), making training and evaluation harder. Falcon-ASR was trained on Emirati, MSA, other Gulf and Arabic dialects, and English, aiming to transcribe everyday speech, including dialectal forms and language switching. Training data included background noise, overlapping speech, music, room reverberation, and telephony effects, plus variations in speed and pitch; the same augmentation was applied to Emirati recordings to prepare the model for meetings, calls, and other everyday conditions.

For Arabic benchmarking, TII followed the protocol of the Open Universal Arabic ASR Leaderboard maintained by the ELM Research Center, which ranks systems by equal-weight average WER across six test sets and also reports CER, with lower values better. Falcon-ASR was evaluated on the same six benchmarks using the leaderboard’s pinned manifests. It achieved an average WER of 20.92%, compared with the best published result of 23.17% in the leaderboard snapshot checked on 30 September 2026—an improvement of 2.25 percentage points. Competitor figures are the published leaderboard averages from that snapshot.

For Emirati speech, TII complemented public data (Casablanca includes a UAE subset) with an internal evaluation of additional Emirati and Gulf speech using held-out recordings and human-validated transcripts. In that internal evaluation, Falcon-ASR recorded 22.73% WER and 10.19% CER, the lowest among the compared systems and 4.07 percentage points below Qwen3-Omni, the next best result. The article states these results show improved transcription accuracy at both the word and character level.

The model also transcribes English with the same weights, achieving a mean WER of 5.74% on the seven public English test sets used by the Hugging Face Open ASR Leaderboard. French, Spanish, and Portuguese are supported as well. Falcon-ASR builds on TII’s Falcon3-Audio work, described in the paper “Competitive Audio-Language Models with Data-Efficient Single-Stage Training on Public Data.” A Hugging Face Demo Space is available for trying the model with your own recordings; API access and native applications are planned. The acknowledgments thank the Falcon-Emirati team and Mikhail Lubinets.