The Technology Innovation Institute in Abu Dhabi has released Falcon-ASR, a 1.6-billion-parameter speech recognition model built around Arabic and in particular the Emirati dialect. It also handles English, French, Spanish and Portuguese from the same weights, with no language flag to set.
Dialect is the reason this is interesting. Most Arabic speech systems are trained and measured on Modern Standard Arabic, which is not what people speak, and transcribed recordings of Gulf dialects are scarce. TII reports an average word error rate of 20.92% across the six test sets of a public Arabic leaderboard, against 23.17% for the best published result in the snapshot it used, checked at the end of September. On English it reports a mean 5.74% across seven public test sets.
The Emirati figure, 22.73% word error and 10.19% character error, comes from TII’s own internal evaluation rather than any public benchmark, which the institute states plainly. That is roughly four points ahead of the next best system it tested. Public data covers the UAE only partly, which is why the internal set exists, but it also means nobody outside TII can reproduce the headline dialect claim.
No licence is named in the announcement, and API access and applications are described as planned. For now there is a demo that takes clips of up to a minute.