TII's Falcon-ASR Hits 20.92% Arabic WER, Beats Larger Rivals

The Technology Innovation Institute (TII) in Abu Dhabi has released Falcon-ASR, a 1.6B-parameter Arabic speech recognition model that posts a 20.92% average word error rate across 6 Arabic test sets — 2.25 points better than the best published entry in the leaderboard snapshot TII used for comparison. The model targets the Emirati dialect in particular and also transcribes English, French, Spanish and Portuguese. For developers, the immediate limitation is access: Falcon-ASR currently runs only as a hosted Hugging Face demo, with no published weights, license, or API.

The stakes are real for Arabic voice products. Dialectal Arabic has far fewer transcribed resources than Modern Standard Arabic, which has historically left regional speech systems trailing their English counterparts. A compact model that outperforms systems several times its size — on TII's own tests at least — signals where the Arabic ASR market is heading.

What is Falcon-ASR?

Falcon-ASR is a speech-to-text model that converts spoken Arabic and four other languages into text using a single set of weights. The model needs no language flag: it detects the spoken language and outputs text in that language automatically.

It also returns word-level timestamps, linking each word to its position in the audio. That matters for subtitling, searchable recordings and meeting notes — applications where alignment between text and audio is a hard requirement.

The model builds on TII's Falcon3-Audio research, described in a paper on arXiv. That work covers audio-language models built on instruction-tuned LLMs and Whisper encoders, trained on under 30K hours of public audio. TII has not published Falcon-ASR's exact architecture or training data size.

How does Falcon-ASR handle Arabic dialects?

TII trained Falcon-ASR on Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English. The stated objective is everyday speech, including dialect forms and language switching — the messy reality of how people actually talk in the region.

The training pipeline leaned heavily on data augmentation. TII added background noise, overlapping speech, music, room reverberation and telephony effects to the training audio. Speed and pitch were varied as well. The same treatment was applied to Emirati recordings, presumably to harden the model against real-world recording conditions.

How does Falcon-ASR score on Arabic benchmarks?

TII followed the protocol of the Open Universal Arabic ASR Leaderboard, run by the ELM Research Center. That benchmark ranks systems by equal-weight average WER over 6 test sets, and TII used the leaderboard's pinned manifests. Competitor numbers 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

Lower is better. One caveat: Falcon-ASR's figure comes from TII's own run, not a leaderboard entry. The Audar model card lists its own average CER as 9.20%.

The pattern is notable. A 1.6B model from TII outperforms a 7B omniASR system by more than 7 WER points in these comparisons. Parameter count alone is not deciding the Arabic ASR race — training data quality and dialect coverage are.

How does Falcon-ASR perform on Emirati speech?

TII built an internal Emirati and Gulf test set from held-out recordings with human-validated transcripts. It complements the UAE subset of the public Casablanca dataset from UBC-NLP.

Model	Params	WER (%)	CER (%)
Falcon-ASR	1.6B	22.73	10.19
Qwen3-Omni-30B-A3B-Instruct	30B (3B active)	26.80	12.72
Audar-ASR-V1-Turbo	2.35B	27.89	13.75
Cohere Transcribe Arabic	2.0B	31.05	18.07
Qwen3-ASR-1.7B	2.0B	31.52	13.35
Audar-ASR-V1-Flash	0.78B	32.87	15.36

Falcon-ASR leads by 4.07 WER points over Qwen3-Omni, a model roughly 19 times its total size. Two caveats apply. This test set is private, so the results cannot be independently reproduced. And Audar reports 19.4% WER on a different Emirati set called Mixat, which means that vendor's Emirati numbers are not directly comparable to TII's.

How good is Falcon-ASR at English?

On the 7 English test sets used by the Hugging Face Open ASR Leaderboard, TII reports a 5.74% mean WER for Falcon-ASR. Per-set scores range from 1.75% on LibriSpeech clean to 11.86% on Earnings-22, its weakest English result.

For context, Qwen3-ASR-1.7B lists 5.76% mean WER on that same leaderboard. The two models are effectively tied on English. French, Spanish and Portuguese are supported, but TII published no scores for them.

Falcon-ASR vs the closest Arabic ASR models

The competitive picture becomes sharper when the models are placed side by side.

• Organization: TII (UAE) vs AudarAI, Cohere, and Alibaba's Qwen team.
• Parameters: Falcon-ASR at 1.6B; Audar-ASR-V1-Turbo at 2.35B; Cohere Transcribe Arabic at 2B; Qwen3-ASR-1.7B at 1.7B (2B on its Hugging Face card).
• License: Not disclosed for Falcon-ASR; AudarAI Community License v1.0 for Audar; Apache 2.0 for both Cohere and Qwen.
• Open weights: No for Falcon-ASR. Audar ships BF16, GGUF and W4A16 variants; Cohere and Qwen publish weights too.
• Languages: 5 for Falcon-ASR; 30 (Arabic-first) for Audar; Arabic and English for Cohere; 30 languages plus 22 Chinese dialects for Qwen.
• Deployment: Not disclosed for Falcon-ASR. Audar runs on CPU, GPU and edge via GGUF, with W4A16 on 12 GB+ GPUs; Qwen runs on GPU via Transformers or vLLM.
• Access: Falcon-ASR offers a Hugging Face demo with API and native apps planned; competitors already ship weights plus commercial APIs (Audar API, Cohere API, DashScope).

Falcon-ASR figures come from TII's blog. Audar, Cohere and Qwen leaderboard figures come from the ELM leaderboard as reproduced on the Audar card. Emirati WER numbers for all models come from TII's internal evaluation.

Why does the access gap matter?

The benchmark numbers tell one story. The distribution model tells another. Every direct competitor in the comparison table ships open weights, and two of them carry permissive Apache 2.0 licenses. Falcon-ASR ships a demo that accepts clips up to 60 seconds and 20 MB.

That gap shapes who can actually use the model. A startup building an Emirati voice assistant can download Audar's GGUF files today and run them on a 12 GB GPU. The same developer cannot deploy Falcon-ASR anywhere, because TII has not disclosed hardware requirements, published weights, or announced pricing.

The self-reported nature of the scores compounds the caution. Falcon-ASR's headline 20.92% figure is TII's own run against the leaderboard's test sets, not a leaderboard entry. The Emirati results rest on a private test set. Both may hold up under scrutiny, but nobody outside TII has verified them.

What comes next?

TII says an API and native apps are planned, which would move Falcon-ASR from research demonstration to usable product. The Falcon3-Audio lineage — LLM decoders paired with Whisper encoders — suggests the architecture has room to scale beyond 1.6B parameters if TII chooses to.

Until weights or an API arrive, Falcon-ASR is a strong benchmark statement rather than a deployable system. Its significance depends on what TII releases next: open weights would pressure Audar and Cohere directly, while a closed API would position Falcon-ASR as a regional service competing on Emirati-dialect accuracy that rivals have not matched.