TII says its 1.6B Falcon-ASR beats the best published Arabic word error rate by 2.25 points

The 20.92% figure comes from a published protocol. The Open Universal Arabic ASR Leaderboard, maintained by the ELM Research Center, ranks systems by the equal-weight average WER across six test sets and reports CER beside it ^[7]. TII ran Falcon-ASR on the leaderboard's pinned manifests and compared the result with competitors' published averages as checked on 30 September 2026 ^[9]. So Falcon-ASR's score is TII's own run, and the competitors' scores are the leaderboard's ^[9]. The 23.17% it beat belongs to Audar-ASR-V1-Turbo ^[6]. A 2.25-point lead ^[5] is about 9.7% fewer word errors in relative terms ^[22]. TII puts the matching average CER at 8.79% ^[20].

Equal weighting decides how far that number travels. Each of the six sets counts the same in the average ^[7]. A model can therefore lead the average while trailing on the one set that sounds most like a given product's traffic. For 20.92% to hold in production, the production audio has to resemble the six-set mix. TII wrote in the announcement: "A model that handles a formal news broadcast may still struggle with a conversation in Emirati or with speech recorded over a phone line." ^[17]

The Emirati result has a different provenance. TII built that evaluation from held-out Emirati and Gulf recordings with human-validated transcripts ^[1]. Qwen3-Omni came second at 26.80% WER ^[2], 4.07 points behind ^[3], a relative gap of about 15% ^[23]. RuntimeWire wrote that the gap "remains a result from TII's own evaluation and the systems TII chose to compare" ^[8]. A public cross-check exists. The Casablanca evaluation data includes a UAE subset ^[21].

The size comparison needs a footnote. RuntimeWire lists Qwen3-Omni at 30 billion parameters with 3 billion active ^[2]. Against Falcon-ASR's 1.6 billion ^[4], it has roughly 19 times the total but under twice the active count ^[24].

I think the training recipe is aimed at the right audio. TII says it trained on Emirati, MSA, other Gulf and Arabic dialects, and English, with the aim of transcribing everyday speech including switches between languages ^[14]. It added background noise, overlapping speech, music, room reverberation, telephony effects, and speed and pitch changes, and applied the same treatment to the Emirati recordings to cover meetings and calls ^[15]. On English, it reports a 5.74% mean WER across the seven public sets used by the Hugging Face Open ASR Leaderboard ^[13]. The model extends TII's Falcon3-Audio work ^[16]. Two of the four credited authors, Abdul Muneer and Ludovick Lepauloux, co-authored that earlier research ^[18].

Beyond the demo and the planned API ^[11], the announcement does not mention downloadable weights or a license. RuntimeWire called the demo "the announced route to test it rather than evidence of a production service" ^[19].