The Technology Innovation Institute in Abu Dhabi has released Falcon-ASR, a 1.6-billion-parameter speech recognition model built primarily for Arabic, with particular attention to the Emirati dialect.
On the Open Universal Arabic ASR Leaderboard's six test sets, Falcon-ASR achieved an average word error rate of 20.92%, against a best published result of 23.17% in the snapshot TII used — a 2.25 percentage point improvement. On TII's internal Emirati evaluation, the model recorded 22.73% WER and 10.19% character error rate, the lowest among the systems compared, beating the next-best result (Qwen3-Omni) by 4.07 percentage points on WER.
The model was trained on Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English, with training data that included background noise, overlapping speech, music, reverberation and telephony effects to reflect real-world recording conditions such as calls and meetings.
Beyond Arabic, Falcon-ASR transcribes English, French, Spanish and Portuguese using the same model weights, without needing a language flag specified in advance. On the Hugging Face Open ASR Leaderboard's seven public English test sets, it posted a mean WER of 5.74%. The model also supports word-level timestamps, linking each transcribed word to its position in the audio.
Falcon-ASR builds on TII's earlier Falcon3-Audio work. A demo is available on Hugging Face for testing with user-supplied recordings; API access and native applications are described as planned but not yet available.
Top comments (0)