Abu Dhabi's Technology Innovation Institute (TII) has extended the Falcon family from text into speech: Falcon-ASR, an open 1.6B-parameter speech recognition model focused on Arabic, with particular attention to the Emirati dialect among Gulf varieties. The same weights also handle English, French, Spanish, and Portuguese — five languages sharing one set of weights with no language flag required. The team published full evaluations in a Hugging Face blog post and opened an online demo.

Arabic leaderboard: 1.6B posts 20.92% average WER

The evaluation follows the protocol of the Open Universal Arabic ASR Leaderboard (maintained by the ELM Research Center): six Arabic test sets, equal-weight average WER, with CER also reported. Falcon-ASR scores an average WER of 20.92% and CER of 8.79%, which is 2.25 percentage points below the best published result in the leaderboard snapshot used (23.17%). The comparison from the official blog:

Model Parameters Avg WER Avg CER
Falcon-ASR 1.6B 20.92 8.79
Audar-ASR-V1-Turbo 2.35B 23.17 9.23
Cohere Transcribe Arabic (07-2026) 2.0B 25.87 11.80
omniASR LLM 7B 7.0B 28.32 12.52

Competitor figures are the published leaderboard averages verified on 30 September 2026 — not numbers TII ran itself.

Dialects are the battleground: 22.73% on internal Emirati eval

The hard part of Arabic ASR has never been formal news-broadcast speech; it is dialect. Dialectal Arabic has far fewer transcribed resources than Modern Standard Arabic, which makes both training and evaluation harder. TII ran an internal evaluation on held-out Emirati and Gulf recordings with human-validated transcripts: Falcon-ASR achieved 22.73% WER and 10.19% CER, the lowest word and character error rates among the systems compared — 4.07 points below Qwen3-Omni, the next best.

Training data deliberately mixed in background noise, overlapping speech, music, room reverberation, telephony effects, and variations in speed and pitch, with the same treatment applied to Emirati recordings — aimed squarely at real meetings and phone calls.

English and other languages: one set of weights, no language flag

On the English side, evaluated across the seven public test sets used by the Hugging Face Open ASR Leaderboard, Falcon-ASR averages 5.74% WER, spanning LibriSpeech clean at 1.75% to Earnings-22 at 11.86%. French, Spanish, and Portuguese are supported by the same weights, with output following the language spoken. Word-level timestamps link each transcribed word to its position in the audio — a direct win for captioning and retrieval applications.

Architecturally, Falcon-ASR builds on TII's Falcon3-Audio work; the training approach is described in their paper (arXiv:2509.07526).

So what

The Falcon family has been stacking up the Arabic stack over the past two months: Falcon-Emirati (dialect LLM), Falcon OCR Arabic, and now Falcon-ASR. A 1.6B model is not large by itself; but "dialect data is scarce" is the shared condition of every low-resource language. TII's playbook — dialect-first training plus all-conditions data augmentation, then self-verifying under a public leaderboard protocol — is a copyable recipe for teams working on Chinese dialects or Southeast Asian languages that likewise lack transcription resources. Weights are open; API access and native applications are planned.

Reference: Introducing Falcon ASR (TII official announcement)