TII Ships Arabic ASR With Emirati Focus
Technology Innovation Institute in Abu Dhabi just released Falcon-ASR, a 1.6 billion parameter automatic speech recognition model trained specifically for Arabic with particular attention to Emirati dialect. The model also supports English, French, Spanish, and Portuguese — all from the same weights, no language flag required.
The headline number: 20.92% average word error rate across six Arabic test sets, compared to the best published result of 23.17% on the Open Universal Arabic ASR Leaderboard snapshot they evaluated against (checked September 30, 2026). That's a 2.25 percentage point improvement on the public benchmark.
But the more interesting story is in the Emirati evaluation, where dialectal Arabic meets real-world recording conditions.
The Dialect Problem in Arabic ASR
Arabic ASR has always had a ground-truth problem. Modern Standard Arabic has relatively abundant transcribed resources. Dialectal Arabic — the stuff people actually speak in everyday conversation — has far less. A model that handles formal news broadcasts can still fall apart on phone calls or meetings in Gulf dialects.
Falcon-ASR was trained on Emirati, MSA, other Gulf and Arabic dialects, plus English. The stated goal is transcribing "the words people use in everyday speech, including dialectal forms and changes between languages." That's the right framing: conversational Arabic isn't clean MSA, and code-switching is common.
On TII's internal Emirati evaluation — held-out recordings with human-validated transcripts — Falcon-ASR hit 22.73% WER and 10.19% CER (character error rate). For context, the next-best system they compared was Qwen3-Omni at 26.80% WER. That's a 4.07 percentage point gap.
Benchmark Architecture
The public Arabic evaluation follows the protocol from the Open Universal Arabic ASR Leaderboard, maintained by ELM Research Center. Six test sets, equal-weight averaging for WER and CER. Lower is better on both metrics.
Falcon-ASR's 20.92% average WER came from evaluating on the same six benchmarks using the leaderboard's pinned manifests. The comparison is apples-to-apples: same test sets, same evaluation protocol, published competitor results from the September 2026 snapshot.
The internal Emirati evaluation adds coverage beyond what's in the public Casablanca dataset's UAE subset. TII is essentially saying: the public benchmarks show we're competitive, but here's additional held-out Emirati data where the gap is larger.
Training for Acoustic Chaos
One detail that jumps out: TII augmented training data with background noise, overlapping speech, music, room reverberation, telephony effects, speed variations, and pitch shifts. They applied the same treatment specifically to Emirati recordings.
This matters because conversational ASR lives or dies on robustness to recording conditions. Meetings, phone calls, street recordings — these aren't studio audio. Training for acoustic variability is table stakes, but explicitly calling out the augmentation pipeline for dialectal data suggests TII knows where the failure modes are.
The model also supports word-level timestamps, linking each transcribed word to its audio position. Useful for downstream applications that need alignment, not just a blob of text.
English and Multilingual Performance
Falcon-ASR hit 5.74% mean WER on the seven public English test sets used by the Hugging Face Open ASR Leaderboard. That's respectable but not groundbreaking for English ASR in 2026.
The multilingual setup is cleaner than typical: French, Spanish, Portuguese, English, and Arabic all use identical weights. No language identifier needed at inference. The model just outputs a transcript in whatever language it hears. For a 1.6B-parameter model, that's efficient parameter usage.
Foundation: Falcon3-Audio
Falcon-ASR builds on TII's Falcon3-Audio work. The architecture and training approach are detailed in their paper "Competitive Audio-Language Models with Data-Efficient Single-Stage Training on Public Data". Single-stage training on public data is doing a lot of work in that title — it's a claim about training efficiency and data accessibility.
TII also credits the Falcon-Emirati team for Arabic foundation model support. The institutional muscle at TII clearly involves multiple model families (Falcon LLMs, Falcon-Emirati for Arabic NLP, now Falcon-ASR for speech), all under the Falcon umbrella.
What's Missing
The blog post doesn't specify training data sources beyond "Emirati, MSA, other Gulf and Arabic dialects, and English." No dataset names, no hour counts, no breakdown of dialectal distribution. For a model claiming Emirati focus, knowing how much Emirati audio went into training would contextualize the performance claims.
Likewise, the internal Emirati evaluation is described but not released. We get aggregate WER and CER, but no test-set details, no domain breakdown (meetings vs. calls vs. other), no speaker demographics. That's a missed opportunity for transparency, especially when the public benchmark advantage is smaller than the internal one.
Finally: no latency numbers, no compute requirements for inference, no discussion of streaming vs. batch. For production deployment, those details matter as much as WER.
The Regional ASR Bet
Falcon-ASR is interesting less for the absolute benchmark numbers — plenty of English ASR models hit similar or better WER — and more for what it represents: a well-resourced bet on dialectal Arabic ASR from a UAE-based research institute.
Emirati Arabic isn't a massive commercial market compared to English or Mandarin. But if you're TII in Abu Dhabi, building state-of-the-art Emirati ASR is both a technical milestone and a strategic asset. This is speech technology that handles the language people actually speak in the region, trained and evaluated with local dialects front and center.
The model is available now on Hugging Face with a demo space. API access and native applications are "planned," which usually means "not yet." But the weights are out there, and 1.6B parameters is small enough to run locally if you need it.
For Arabic ASR practitioners: Falcon-ASR just moved the goalposts on the public leaderboard. For everyone else: this is what regional AI investment looks like when it's executed well.