The Pulse
TII Releases Falcon-ASR, a 1.6B Arabic Speech Model
Abu Dhabi’s Technology Innovation Institute launched Falcon-ASR, a 1.6-billion-parameter speech-recognition model for Emirati and Modern Standard Arabic, plus four other languages. TII reports that it scored 20.92 average word error rate ac
AI.info Team · Written with AI ·
TII’s new Falcon-ASR has 1.6 billion parameters and a reported average word error rate of 20.92 across Arabic test sets. The Technology Innovation Institute announced the speech-recognition model in Abu Dhabi on October 6, alongside Falcon-Emirati and Falcon-OCR-Arabic. Falcon-ASR transcribes Emirati Arabic and Modern Standard Arabic, as well as English, French, Spanish and Portuguese.
Falcon-ASR targets spoken Arabic and five languages
TII describes Falcon-ASR as a compact model for turning speech into text. Its language list spans a formal Arabic variety and a regional dialect, alongside four widely used European languages. That combination is aimed at multilingual settings in which people may switch languages or speak in dialect rather than formal Arabic.
The institute says the model is designed for audio with background noise, overlapping speech, room reverberation and telephone-quality sound. It also produces word-level timestamps, which can align each transcribed word with its point in an audio or video recording. TII lists subtitling, searchable recordings, meeting and interview transcription, accessibility tools and voice-enabled services as possible uses.
TII reports lower error rates than listed rivals
On the Open Universal Arabic ASR Leaderboard comparison shown on TII’s model page, Falcon-ASR posts an average word error rate of 20.92 and character error rate of 8.79. The same table lists Audar-ASR-V1-turbo at 23.17 WER, Cohere Transcribe Arabic at 25.87 and omniASR LLM 7B at 28.32. Lower scores indicate fewer transcription errors, according to TII.
For Emirati speech, TII reports a 22.73 WER and 10.19 CER. Its internal comparison lists Qwen3-Omni-30B-A3B-Instruct at 26.80 WER and 12.72 CER; the model has 30 billion parameters overall, with 3 billion active. These are figures from TII’s published comparisons, including an internal benchmark, not an independent evaluation of the release.
Dialect support addresses a gap in speech systems
In its launch announcement, TII says Arabic dialects remain underrepresented in AI training data. A system can recognize the words in an Emirati sentence, the institute argues, yet miss the intended meaning because dialect expressions carry local rhythms and cultural references. Falcon-ASR focuses on transcription, while the separately announced Falcon-Emirati model is built for generating and understanding Emirati Arabic.
Sovereign AI capability must reflect the language used in daily life. Emirati Arabic carries distinctive expressions, cultural references and ways of communicating.
Dr. Najwa Aaraj, chief executive officer of TII
The product page links to a Falcon-ASR demo, giving users a way to try the system. The announcement and model page describe the model’s capabilities and benchmark results, but do not provide the evaluation audio, full testing protocol or enough detail to independently reproduce the comparisons. Those details matter when comparing systems tested on different speech, dialects and recording conditions.
Benchmark conditions will shape the comparison
Word error rate measures transcription mistakes against a reference transcript; it does not by itself show how performance varies by dialect, accent, background noise or recording device. TII’s figures make Falcon-ASR’s scores legible beside named competitors, but the product page does not describe the test-set composition in enough detail to judge every difference. The Emirati result is explicitly from TII’s internal benchmark.
For developers, the immediate practical questions are how Falcon-ASR performs on their own recordings and what access options TII will provide beyond the demo. The announcement identifies likely applications, but does not set out a deployment license or a reproducible benchmark package. Which dialects and audio conditions drive the reported scores, and can outside teams test the same samples?