Urgent.News

What's breaking now, across hundreds of outlets.

AI

Introducing Falcon ASR

Falcon-ASR is a new speech recognition model developed by the Technology Innovation Institute (TII) in Abu Dhabi, designed specifically for Arabic language transcription. The model, which contains 1.6 billion parameters, is focused on the Emirati dialect, but also supports English, French, Spanish, and Portuguese. In evaluations, Falcon-ASR achieved an average word error rate (WER) of 20.92% across six Arabic test sets, compared to the best published result of 23.17%.

On an internal evaluation of Emirati dialect speech, Falcon-ASR had the lowest WER and character error rate (CER), with 22.73% and 10.19%, respectively. The model also transcribes English, French, Spanish, and Portuguese using the same weights, without requiring a language flag. Falcon-ASR was trained on a diverse range of Arabic dialects, Modern Standard Arabic (MSA), and Gulf languages, in addition to English.

The goal was to transcribe everyday speech, including dialectal forms and language changes. The model's performance was evaluated using the Open Universal Arabic ASR Leaderboard, which ranks systems by the equal-weight average WER across six test sets. Falcon-ASR's average WER is 2.25 percentage points better than the best published result.

The model also included background noise, overlapping speech, music, room reverberation, and telephony effects during evaluation. Falcon-ASR builds on the company's Falcon3-Audio architecture, which is described in detail in a previous research paper. A demo of Falcon-ASR is available on Hugging Face, with plans for API access and native applications.

Rewritten by urgent.news; may contain errors — check the original.

Read the original at huggingface.co →

Report a problem with this story

More in AI

More from Wednesday 7 October →