TII says its 1.6B Falcon-ASR beats larger models on its Emirati test

Why it matters

TII is extending its Falcon research into speech tools built around Emirati Arabic, a dialect its team says is underrepresented in transcription data. TII reports a benchmark lead for Falcon-ASR, but broader testing and deployment will determine its practical value.

The Technology Innovation Institute says its new 1.6-billion-parameter Falcon-ASR model transcribed Emirati speech more accurately than the larger systems it tested, including a 30-billion-parameter multimodal model. The Abu Dhabi research institute announced the model on October 6th as part of a three-model release focused on Arabic language and culture; Hugging Face published the technical results on October 7th.

The work is credited to Abdul Muneer, Ludovick Lepauloux, Rishabh Saraf and Shamsa Hamad. TII does not present Falcon-ASR as a startup or name a conventional founder. Muneer and Lepauloux had also co-authored TII's earlier Falcon3-Audio research, placing the speech recognizer in a continuing research effort rather than a one-off product announcement.

Falcon-ASR builds on Falcon3-Audio and narrows the task to speech transcription, with particular emphasis on Arabic dialects that have fewer transcribed training resources than Modern Standard Arabic. TII says it trained the model on Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English. TII also says training included background noise, overlapping speakers, music, room reverberation, telephone effects, and changes in speaking speed and pitch.

A benchmark lead with boundaries

Across six Arabic test sets, TII reports an average word error rate of 20.92% and character error rate of 8.79%. Word error rate measures the share of words that are incorrectly transcribed; lower is better. The institute compares its score with 23.17% for Audar-ASR-V1-Turbo, the best published result in the leaderboard snapshot it used. Competitor averages were checked on September 30th, 2026, using the same pinned test manifests, according to the announcement.

The larger result for the model's central use case comes from TII's internal Emirati evaluation: 22.73% WER and 10.19% CER. In that comparison, the next-lowest WER was 26.80% for Qwen3-Omni, a model listed at 30 billion parameters, with 3 billion active. The gap is notable, but it remains a result from TII's own evaluation and the systems TII chose to compare. The public leaderboard and internal test answer different questions, so the internal number should not be read as a universal ranking of Arabic speech systems.

RuntimeWire reported separately that TII's Falcon-Emirati model scored 84.83% on a benchmark of Emirati dialect and cultural knowledge. Falcon-ASR focuses on turning spoken Emirati Arabic into text that can be searched, captioned or used in other software. The October 6th announcement also covered text generation and included Falcon-OCR-Arabic, which reads Arabic text from images and documents. The release spanned speech recognition, text generation and OCR.

From research result to usable tool

Falcon-ASR supports Modern Standard Arabic, Emirati Arabic, English, French, Spanish and Portuguese, using the same model weights without requiring users to specify the spoken language first, TII says. It can also attach word-level timestamps, which connect each word in a transcript to its position in the recording. Those features are aimed at workflows such as meeting transcription, subtitling, searchable recordings and accessibility tools.

TII reports a mean WER of 5.74% across seven public English test sets, but the Arabic results are the sharper differentiator in this launch. For Emirati Arabic, the institute says the model's dialect coverage is intended to handle everyday speech rather than only formal broadcasts. Transcription quality can change with regional vocabulary, speaker, recording conditions and code-switching.

For now, users can try the model through the Hugging Face demo. TII says API access and native applications are planned, making the demo the announced route to test it rather than evidence of a production service. The launch reports a benchmark lead, while real-world adoption will depend on performance across varied speakers and recordings as well as how TII packages the model for deployment.