Introducing Falcon ASR

Hugging Face has introduced Falcon-ASR, a 1.6 billion parameter speech recognition model built at the Technology Innovation Institute in Abu Dhabi. The pitch is straightforward: better Arabic transcription, especially for Emirati speech, without giving up support for English, French, Spanish or Portuguese.

The numbers are the hook. On six Arabic test sets, Falcon-ASR posted a 20.92% average word error rate. The best published result in the leaderboard snapshot the team used was 23.17%. In the model’s internal Emirati evaluation, it also came out on top, with the lowest word and character error rates among the systems compared.

That Emirati focus matters because Arabic speech is messy in the real world. A model can do fine on formal news audio and still stumble on dialect, on phone recordings, or on speakers who move between Arabic and English. Hugging Face says Falcon-ASR was trained on Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, plus English, with noisy and difficult conditions in the mix: overlapping speech, music, reverberation, telephony, speed changes, and pitch changes.

The model is not just about Arabic, either. On seven public English test sets used by the Hugging Face Open ASR Leaderboard, it reached a 5.74% mean word error rate. And the same weights handle all five supported languages without a language flag, which is cleaner than the usual menu of settings and switches.

There’s also a small practical touch that should save people some pain: word-level timestamps, so each transcribed word is linked to its place in the audio. Falcon-ASR is available in a Hugging Face Demo Space now, while API access and native apps are still planned.