TII releases Falcon-ASR, a 1.6B speech recognition model focused on Arabic and Emirati dialect
In short: The Technology Innovation Institute (TII) has released Falcon-ASR, a 1.6 billion parameter automatic speech recognition model built primarily for Arabic, with particular focus on the Emirati dialect. It also transcribes English, French, Spanish and Portuguese using the same model weights. TII reports it beats the best published leaderboard result on a standard Arabic benchmark and leads on an internal Emirati evaluation, while also including word-level timestamps.
This summary was generated automatically by AI from Hugging Face's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 11.6B parameter model supporting Arabic (incl. Emirati, Gulf dialects, MSA), English, French, Spanish, Portuguese with one set of weights, no language flag needed
- 2Achieved 20.92% average WER on six Arabic test sets from the Open Universal Arabic ASR Leaderboard, vs. 23.17% for the next best published result (Audar-ASR-V1-Turbo)
- 3On an internal Emirati evaluation scored 22.73% WER / 10.19% CER, ahead of Qwen3-Omni-30B-A3B-Instruct (26.80% WER) and other compared systems
- 4Achieved 5.74% mean WER across seven public English test sets (Hugging Face Open ASR Leaderboard)
- 5Supports word-level timestamps linking transcribed words to audio position
- 6Trained with background noise, overlapping speech, reverberation and telephony effects to improve robustness in calls and meetings
- 7Available now via a Hugging Face demo; API access and native apps are planned but not yet available
| Parameter | Before | Now |
|---|---|---|
| Parameters | Not specified | 1.6B |
| Languages | Not specified | Arabic (incl. Emirati dialect), English, French, Spanish, Portuguese |
| Arabic avg WER | 23.17% (best published) | 20.92% |
| Emirati WER (internal eval) | 26.80% (next best, Qwen3-Omni) | 22.73% |
| English avg WER (7 public sets) | Not specified | 5.74% |
| Timestamps | Not specified | Word-level timestamps supported |
| API access | Not available | Planned (not yet available) |
Why it matters
It shows continued progress in open ASR models specifically tuned for underrepresented dialects and noisy/telephony conditions, which matters for any team building call transcription or voice analytics in challenging acoustic environments, even if the language coverage here doesn't include Lithuanian.
What it means for AI agents and contact centers
This model doesn't currently cover Lithuanian, so it isn't directly usable for call transcription in that market, but its explicit training on telephony effects, noise and overlapping speech, plus word-level timestamps, is a useful reference design when evaluating or benchmarking STT engines for contact-center pipelines; worth watching for future language expansions and the planned API.
🧪 Worth Testing
Testing the demo could help benchmark robustness techniques (noise, telephony artifacts, timestamps) against the STT engines currently used in call pipelines, even without Lithuanian support.
Sources
- Hugging FaceOfficialPrimary sourceOriginal article →„Introducing Falcon ASR“7 Oct 2026, 16:21
- Published by source
- 7 Oct 2026, 16:21
- Found by our system
- 7 Oct 2026, 16:36
- Summary generated
- 7 Oct 2026, 16:37
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy