TII Releases Falcon-ASR: 1.6B Arabic Speech Model Hits 20.92% WER

The Technology Innovation Institute in Abu Dhabi has released Falcon-ASR, a 1.6 billion parameter speech-to-text model built around Arabic, with a focus on the Emirati dialect, plus English, French, Spanish, and Portuguese on one shared set of weights. In TII's evaluation it posts a 20.92% average word error rate across six Arabic test sets, 2.25 points ahead of the next published entry in its comparison snapshot.

For anyone building Arabic voice products, dialectal speech has always been the hard part. Modern Standard Arabic gets the data. Everyday Emirati and Gulf speech does not. A compact model that claims to lead there is news. The access catch is also news: today it is a Hugging Face demo only.

We want this kind of work to succeed. Arabic dialects are underserved relative to English ASR, and a 1.6B model that beats larger systems on dialectal speech is exactly the direction useful AI should go — capability for languages and communities that the big labs treat as secondary. Word-level timestamps and no language flag required make it practical for subtitling, searchable recordings, and meeting notes.

I'm watching the gap between the leaderboard and what developers can actually ship. Falcon-ASR's 20.92% Arabic average WER and 8.79% character error rate beat Audar-ASR-V1-Turbo at 23.17%, Cohere Transcribe Arabic at 25.87%, and a 7B omniASR at 28.32% in TII's table. On TII's internal Emirati and Gulf set it posts 22.73% WER, 4.07 points ahead of Qwen3-Omni-30B-A3B-Instruct. English mean WER on the seven Hugging Face Open ASR Leaderboard sets is 5.74%, roughly matching Qwen3-ASR-1.7B's 5.76%. Those are strong numbers for a compact model.

They are also almost entirely self-reported. The Arabic average is TII's own run against the Open Universal Arabic ASR Leaderboard protocol, not a published leaderboard entry. The Emirati set is private and cannot be reproduced. MarkTechPost could not find Falcon-ASR's exact architecture or training data size; the model builds on Falcon3-Audio research that used under 30,000 hours of public audio, but that is not a Falcon-ASR disclosure. French, Spanish, and Portuguese have support claims and no published scores.

The distribution story undercuts the benchmark story. No open weights. No disclosed license. No API yet — only plans. Audar, Cohere, and Qwen already ship weights. A hosted demo capped at sixty seconds and twenty megabytes is a preview, not a product. If the scores hold under independent replication, Falcon-ASR deserves to be a default for Emirati and Gulf ASR. Until weights and a license land, competitors with slightly worse numbers and actual downloadable models will win the installs.

Our read: celebrate the Emirati focus and the compact size, demand third-party runs on the public Arabic leaderboard, and treat "API planned" as unfinished business. Useful dialect ASR only counts when builders can run it.

TII followed the Open Universal Arabic ASR Leaderboard protocol from the ELM Research Center, averaging word error rate equally across six test sets with pinned manifests. Competitor averages were checked as of September 30, 2026. Training included Emirati, MSA, other Gulf and Arabic dialects, and English, with augmentation for noise, overlap, music, reverberation, telephony, speed, and pitch.