The Technology Innovation Institute (TII) in Abu Dhabi introduced Falcon-ASR, a 1.6B-parameter speech model with an Emirati dialect focus. It transcribes Arabic plus English, French, Spanish, and Portuguese, and returns word-level timestamps. In TII’s own evaluation, it posts a 20.92% average WER across six Arabic test sets, beating the best published leaderboard entry by 2.25 points. There is a catch for developers: access is limited to a hosted demo for now.
What is Falcon-ASR and what makes it different?
Falcon-ASR is TII’s speech-to-text model that handles Arabic and four other languages with a single set of weights. It needs no language flag and outputs text in the language spoken in the audio. The model also returns word-level timestamps, linking each token to its position in the clip.
For you, that enables accurate subtitling, searchable transcripts, and meeting notes. Word-level alignment speeds navigation through long recordings and simplifies editing. It helps media teams, contact centers, and internal content archives.
The model supports five languages: Arabic (with an Emirati emphasis), English, French, Spanish, and Portuguese. One set of weights serves all five, so there is no mode switching or language signaling. That simplifies integration where code-switching appears.
The model size is 1.6 billion parameters. The maximum audio context is not disclosed. The public demo accepts clips up to 60 seconds and 20 MB, enabling quick trials without extra setup.
How does Falcon-ASR handle Arabic dialects?
The short answer: it was trained on Emirati, MSA, other Gulf and Arabic dialects, plus English. The objective targets everyday speech, including dialectal forms and language switching that occurs in real use.
Data preparation used heavy augmentation: background noise, overlapping speakers, music, room reverberation, and telephony effects. Speed and pitch were varied too. The same treatment was applied to Emirati recordings to build robustness.
Falcon-ASR builds on TII’s Falcon3-Audio work. That research describes audio-language models based on instruction-tuned LLMs and Whisper encoders, trained on under 30K hours of public audio. This offers context on the approach to unifying speech and language tasks.
At the same time, no exact architecture for Falcon-ASR or training data size is published in the available materials. For you, that means design details and corpus scale remain undisclosed despite the reported results.
How does Falcon-ASR score on Arabic benchmarks?
TII followed the protocol of the Open Universal Arabic ASR Leaderboard by the ELM Research Center. Systems are ranked by an equal-weight average WER over six test sets. The leaderboard’s pinned manifests were used, and competitor scores are published averages as of September 30, 2026.
Within this setup, Falcon-ASR reports 20.92% average WER and 8.79% average CER. That is 2.25 points below Audar-ASR-V1-Turbo at 23.17% WER (9.23% CER in the reproduced table). Cohere Transcribe Arabic stands at 25.87% WER and 11.80% CER, while omniASR LLM 7B scores 28.32% WER and 12.52% CER.
Important note: lower is better. Falcon-ASR’s figure is TII’s own run rather than an official leaderboard entry. Also, the Audar model card lists 9.20% as its average CER, which is worth noting when reading the numbers.
Why should you care? These results suggest a compact 1.6B model can compete with larger Arabic systems on recognition metrics. That is a signal for products where Arabic is the primary interaction language.
20.92% average WER across six Arabic sets — and 2.25 points under Audar-ASR-V1-Turbo’s published reference.
How does Falcon-ASR perform on Emirati speech?
On TII’s internal Emirati and Gulf test set, Falcon-ASR reaches 22.73% WER and 10.19% CER. The set was built from held-out recordings with human-validated transcripts and complements the UAE subset of the public Casablanca dataset.
On that same internal set, Qwen3-Omni-30B-A3B-Instruct reports 26.80% WER and 12.72% CER; Audar-ASR-V1-Turbo shows 27.89% and 13.75%; Cohere Transcribe Arabic posts 31.05% and 18.07%; Qwen3-ASR-1.7B gets 31.52% and 13.35%; and Audar-ASR-V1-Flash comes in at 32.87% and 15.36%.
Falcon-ASR leads Qwen3-Omni by 4.07 WER points in this test. That highlights strength on Emirati speech, which is often underrepresented in public resources.
However, this test set is private and cannot be reproduced externally. Also, Audar reports 19.4% WER on a different Emirati set (Mixat), so the figures are not directly comparable across datasets.
The Emirati evaluation set is private; results are not externally reproducible or directly comparable with other datasets.
How strong is its English — and how can you access it?
Across the seven English test sets used by the Hugging Face Open ASR Leaderboard, TII reports a 5.74% mean WER. Per-set scores range from 1.75% on LibriSpeech clean to 11.86% on Earnings-22. For context, Qwen3-ASR-1.7B lists a 5.76% mean WER on the same leaderboard.
French, Spanish, and Portuguese are also supported, but TII published no scores for those languages. That means the model is presented as multilingual, while evaluations beyond Arabic and English are not provided.
As for access and deployment: today Falcon-ASR is available as a Hugging Face demo. An API and native apps are planned. Model weights and quantized variants are not published, the license is undisclosed, and the runtime platform is not stated.
The demo accepts clips up to 60 seconds and 20 MB. The maximum audio context is not disclosed. That’s convenient for quick quality checks, yet it limits production integration right now.
Bottom line: at best, a compact 1.6B model leading larger Arabic systems in TII’s tests. At worst, no weights, license, or API yet — and all scores are self-reported.
Based on TII Falcon-ASR and the ELM Arabic ASR Leaderboard snapshot.
