Falcon ASR opens multilingual speech recognition focused on Emirati Arabic as a live demo

What is changing?

The Technology Innovation Institute in Abu Dhabi has introduced Falcon ASR, a 1.6-billion-parameter speech recognition model. It emphasizes Emirati Arabic while also covering Modern Standard Arabic and other Gulf dialects. The same weights transcribe English, French, Spanish, and Portuguese without a language flag, and the model can produce word-level timestamps.

In TII's published evaluation, Falcon ASR reached a 20.92% word error rate across the equal-weighted average of six Arabic test sets. The previous best published result in the comparison was 23.17%. TII reported 22.73% on its own Emirati test; that internal evaluation is not an independent benchmark.

The model can currently be tested on the official Hugging Face demo with audio clips of up to 60 seconds. TII says API access and native applications are planned, so today's access is not a production API. Dialect coverage is useful for call, meeting, and captioning teams, but they should measure errors on their own noise, code-switching, and domain-terminology data before deployment.