The Technology Innovation Institute introduced Falcon ASR on October 7, 2026 through its Hugging Face blog. The 1.6-billion-parameter speech-recognition model focuses on Arabic, particularly the Emirati dialect, and also supports English, French, Spanish and Portuguese.
TII reports an average word-error rate of 20.92 percent across six Arabic test sets, compared with a best published result of 23.17 percent in the leaderboard snapshot it used. Lower word-error rates indicate fewer transcription errors in that evaluation. They are not a promise of accuracy on every dialect, speaker or recording.
The institute also reports a 22.73 percent word-error rate on its internal Emirati evaluation and says it had the lowest word and character error rates among the systems compared. This is TII's own evaluation, not an independent reproduction.
Falcon ASR supports word-level timestamps, connecting transcribed words to positions in an audio recording. TII says its training covered Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English, with an emphasis on everyday speech and language switching.
The announcement discusses differences between broadcast speech, conversations and phone recordings. A single benchmark does not establish accuracy across all real-world audio conditions. Users should test their own dialect and recording quality before relying on transcripts.
A Hugging Face demo is linked from the announcement. We have not tested that demo or verified production throughput. The confirmed news is the model announcement and the capabilities and evaluation results reported by its developer.