Introducing Falcon ASR

What it claims

Falcon-ASR is a 1.6B-parameter speech recognition model from TII that targets Emirati Arabic. One set of weights covers Arabic (MSA, Emirati, other Gulf and Arabic dialects), English, French, Spanish and Portuguese. You do not pass a language flag; the model outputs the transcript in whatever language is spoken. The page states the training goal as handling everyday dialectal speech and switches between languages.

How it works

• Base: Falcon-ASR builds on TII's Falcon3-Audio. That model's architecture and training are described in Competitive Audio-Language Models with Data-Efficient Single-Stage Training on Public Data. The blog gives no architecture details of its own.
• Training languages: Emirati, MSA, other Gulf and Arabic dialects, and English.
• Augmentation: background noise, overlapping speech, music, room reverberation, telephony effects, and speed and pitch perturbation. The Emirati recordings got the same treatment. The stated aim is to cover meetings, calls and phone-line audio.

Evidence

• Arabic leaderboard: 20.92% average WER on the six test sets of the Open Universal Arabic ASR Leaderboard (ELM Research Center).
• The score is an equal-weight average, computed with the leaderboard's pinned manifests.
• The baseline is 23.17%, the best published leaderboard average as checked on 30 September 2026.
• TII ran this evaluation itself. The page describes it as a comparison against a snapshot, not as a leaderboard entry.
• Internal Emirati set: 22.73% WER and 10.19% CER, the lowest among the systems TII compared. The next best was Qwen3-Omni, 4.07 points higher on WER.
• The set uses held-out Emirati and Gulf recordings with human-validated transcripts.
• It is meant to complement the public UAE subset of Casablanca.
• English: 5.74% mean WER on the seven English test sets used by the Hugging Face Open ASR Leaderboard.

How to use it

The only access the page describes is a Hugging Face Demo Space for trying your own recordings. API access and native applications are "planned". The page gives no model ID, no download, no inference code and no licence. The model outputs word-level timestamps.

Limits

• No stated path to self-hosting. Nothing on the page indicates downloadable weights, and no licence is given. Check the Hub before planning to deploy or fine-tune.
• Emirati result can't be reproduced. The internal set is not public, and the page gives neither its size nor the full list of compared systems (the comparison tables did not come through in the page text).
• No per-test-set numbers. Arabic results appear only as an average. CER on the six-set leaderboard is mentioned but no figure appears in the page text.
• French, Spanish and Portuguese are untested on the page. They are listed as supported but have no reported numbers. They are also absent from the stated training-language list.
• Code-switching and augmentation are unmeasured. Mixed-language transcription is stated as an aim only, and the augmentation has no ablation behind it.
• Comparison is time-bound. The Arabic comparison reflects one leaderboard snapshot and may be stale.