Falcon ASR

Falcon ASR is a speech-to-text model with 1.6 billion parameters from the Technology Innovation Institute in Abu Dhabi, announced on 7 October 2026. One set of weights covers Arabic, with particular attention to the Emirati dialect, and English, French, Spanish and Portuguese, so callers do not set a language flag. TII reports its own evaluations: an average word error rate of 20.92% across six Arabic test sets under the protocol of the Open Universal Arabic ASR Leaderboard, against a best published 23.17% in the snapshot it used, and 22.73% on an internal Emirati set, which it says is 4.07 points better than Qwen3-Omni, the next best system it compared. On seven English sets from the Hugging Face Open ASR Leaderboard it reports a mean of 5.74%. These are the developer’s own runs, so no score is stored. The post links a public demo, says API access and native apps are planned, and does not state a licence or a weights link, so a buyer cannot yet self-host it or know its terms. No price is stated.

Not yet independently verified: Read on the Technology Innovation Institute’s own post of 7 October 2026; the only dated independent mention found is a Hacker News post on 8 October. No weights link or licence is stated and API access is only planned, so this listing covers the announcement and a demo. The error rates are TII’s own evaluations, so no score is stored. We will update this when it can be confirmed, and remove this note.