🔍 Read the full analysis: Meet Falcon ASR, An AI Tool For Turning Speech Into Text on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
The Technology Innovation Institute in Abu Dhabi has introduced Falcon-ASR, a 1.6-billion-parameter speech recognition model focused on Arabic, including Emirati dialect. TII reports a 20.92% average word error rate across six Arabic test sets and strong results in an internal Emirati evaluation; the model is available through a Hugging Face demo.
The Technology Innovation Institute (TII) in Abu Dhabi has introduced Falcon-ASR, a 1.6-billion-parameter model that converts speech into text, with a particular focus on Arabic and the Emirati dialect, as described in the original analysis. TII reports a 20.92% average word error rate across six Arabic test sets and says the system achieved the lowest error rates among the systems in its internal Emirati comparison; people can try it now through a Hugging Face demo.
TII says Falcon-ASR supports Arabic, English, French, Spanish and Portuguese, using the same model weights for all five languages without requiring users to specify a language flag. The institute says its training included Emirati and Modern Standard Arabic, other Gulf varieties and Arabic dialects, as well as English. Training data also included conditions such as background noise, overlapping speech, music and telephone audio, according to TII.
For the Arabic benchmark, TII reports an equal-weight average word error rate (WER) of 20.92% across the six test sets used by the Open Universal Arabic ASR Leaderboard. In the leaderboard snapshot it checked on September 30, 2026, the best published average was 23.17%, a difference of 2.25 percentage points. Lower WER means fewer word-level transcription errors. TII says it followed the leaderboard protocol and used its pinned manifests.
For Emirati speech, TII reports 22.73% WER and 10.19% character error rate (CER) in an internal evaluation using held-out Emirati and Gulf recordings with human-validated transcripts. The institute says those were the lowest scores among the systems it compared, and that the next-best WER, from Qwen3-Omni, was 4.07 percentage points higher. Falcon-ASR also returns word-level timestamps, linking transcribed words to their positions in the audio.
Dialect Speech in Focus
Falcon-ASR targets a practical weakness in speech recognition: performance can vary between formal Arabic and everyday regional speech. Spoken Arabic differs across communities, and training material for dialects is less available than material for Modern Standard Arabic. TII’s Emirati-specific evaluation is relevant to people building tools for calls, meetings and ordinary recordings, where speakers may use local dialects or switch languages.
Word-level timestamps could also help users find particular phrases in longer recordings and support workflows that need to align text with audio. But benchmark results describe performance on specified test material; they do not establish accuracy for every speaker, accent or recording environment. The results are reported by the model’s developer, and the supplied information does not describe an independent replication.
As an affiliate, we earn on qualifying purchases.
How TII Tested the Model
The six-set Arabic comparison is based on the Open Universal Arabic ASR Leaderboard, maintained by the ELM Research Center. Its reported average gives equal weight to each test set. TII says it compared its score with published leaderboard results in a snapshot checked on September 30, 2026, so the 23.17% reference is tied to that date rather than a permanent ranking.
TII also points to the public Casablanca dataset, which includes a UAE subset. Separately, it reports a 5.74% mean WER on seven public English test sets used by the Hugging Face Open ASR Leaderboard. The institute says Falcon-ASR builds on its earlier Falcon3-Audio work. These evaluations provide different reference points, but the Arabic leaderboard average and internal Emirati test should not be treated as interchangeable.
“Our aim is to transcribe the words people use in everyday speech, including dialectal forms and switches between languages.”
— Technology Innovation Institute
Limits of the Published Results
The available material does not provide Falcon-ASR’s score for each of the six Arabic test sets, or results broken down by dialect, speaker or recording condition. TII has also not specified the size and full composition of its internal Emirati evaluation or listed every system included in that comparison. Those details would help readers judge how broadly the reported scores apply.
The leaderboard figures reflect a snapshot checked on September 30, 2026; later published results may alter the comparison. The figures come from TII, and no independent replication is described in the supplied material. Actual performance on recordings unlike those in the evaluations, including across different accents and environments, remains unestablished.
Demo Access and Planned Releases
Falcon-ASR is available to try through TII’s Hugging Face Demo Space, where users can explore transcription with their own recordings. TII says it plans to offer API access and native applications, but has not announced release dates. Those products could make the model easier to integrate into software and workplace processes once available.
For now, the demo gives users a chance to inspect outputs on particular audio, while the published scores provide evidence only for the named test settings. Further evaluation details, per-dialect results and independent testing would help clarify how well Falcon-ASR performs beyond those sets.
Key Questions
What is Falcon-ASR?
Falcon-ASR is a 1.6-billion-parameter speech recognition model introduced by Abu Dhabi’s Technology Innovation Institute. It converts spoken audio into text and returns word-level timestamps.
Which languages does Falcon-ASR support?
TII says the model supports Arabic, English, French, Spanish and Portuguese. The institute says the same model weights handle all five languages without a required language flag.
How did Falcon-ASR score on Arabic speech?
TII reports an average 20.92% WER across six Arabic leaderboard test sets. It also reports 22.73% WER and 10.19% CER in its internal Emirati evaluation. These are developer-reported results for the stated test settings.
Can people try Falcon-ASR now?
Yes. TII says a Hugging Face demo is available. API access and native applications are planned, but the institute has not provided release dates.
Do the reported scores predict performance on every recording?
No. The scores describe results on specified benchmark and internal evaluation material. The published information does not establish performance for every dialect, speaker, accent or recording condition.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
