TII releases 1.6B Falcon-ASR, claims 20.92% Arabic WER against best listed 23.17%

The Technology Innovation Institute (TII) in Abu Dhabi has released Falcon-ASR, a 1.6B-parameter speech recognition model for Arabic with a focus on the Emirati dialect. According to TII, it reaches a 20.92% average word error rate (WER) across the six test sets of the Open Universal Arabic ASR Leaderboard protocol. The best published result in the snapshot TII used was 23.17%.

All figures below come from TII's own evaluations and have not been independently verified.

Key specs

• Parameters: 1.6B
• Languages: Arabic (Emirati, Modern Standard Arabic, other Gulf and Arabic dialects), English, French, Spanish, Portuguese
• Language handling: all five languages share the same weights, with no language flag required
• Output: transcripts in the spoken language, with word-level timestamps
• Foundation: builds on TII's Falcon3-Audio work
• Access: Hugging Face demo Space; API and native applications are "planned"
• Pricing, context window, training cutoff: not yet disclosed

Arabic benchmark results

The leaderboard is maintained by the ELM Research Center and ranks systems by equal-weight average WER across six test sets. TII says it used the leaderboard's pinned manifests. Competitor figures are published leaderboard averages checked on 30 September 2026.

Model	Params	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

The gap to the runner-up is 2.25 percentage points of WER.

Emirati dialect evaluation

TII supplemented public data, including the UAE subset of Casablanca, with an internal evaluation. It uses held-out Emirati and Gulf recordings with human-validated transcripts. The dataset is not public, so the results cannot be independently reproduced.

Model	Params	WER (%)	CER (%)
Falcon-ASR	1.6B	22.73	10.19
Qwen3-Omni-30B-A3B-Instruct	30.0B (3.0B active)	26.80	12.72
Audar-ASR-V1-Turbo	2.35B	27.89	13.75
Cohere Transcribe Arabic (07-2026)	2.0B	31.05	18.07
Qwen3-ASR-1.7B-hf	2.0B	31.52	13.35
Audar-ASR-V1-Flash	0.78B	32.87	15.36

TII reports a 4.07-point WER advantage over Qwen3-Omni, the next-best system in this comparison.

English performance

On the seven public test sets used by the Hugging Face Open ASR Leaderboard, TII reports a mean WER of 5.74%:

• LibriSpeech clean: 1.75%
• LibriSpeech other: 4.21%
• SPGISpeech: 2.02%
• VoxPopuli: 3.87%
• GigaSpeech: 8.15%
• AMI: 8.33%
• Earnings-22: 11.86%

TII did not publish comparison numbers for other models on English in the announcement. It also did not publish per-language results for French, Spanish or Portuguese.

Training data and conditions

TII says it trained on Emirati, MSA, other Gulf and Arabic dialects, and English. It augmented data with background noise, overlapping speech, music, room reverberation, telephony effects, and speed and pitch variation. The same treatment was applied to Emirati recordings. The architecture and training approach for Falcon3-Audio are described in TII's paper, "Competitive Audio-Language Models with Data-Efficient Single-Stage Training on Public Data." The announcement does not state the size or composition of the Falcon-ASR training set.

What this means

A 1.6B model that beats larger Arabic systems on a public leaderboard protocol, if the numbers hold, shows that dialect-focused data can matter more than parameter count. The result is strongest on the public Arabic benchmark, where the protocol is fixed and the baselines are published. The Emirati lead rests on an internal test set that outside parties cannot check. Treat it as directional until independent evaluations appear.

Three things are missing for production teams: pricing, a stated license, and confirmed weight availability. The announcement points only to a demo, with API access "planned." Until those details arrive, Falcon-ASR is something to evaluate, not yet something to deploy.