The Expertise Innovation Institute (TII) in Abu Dhabi has launched Falcon-ASR, a 1.6B parameter Arabic speech recognition mannequin. It focuses on the Emirati dialect and likewise transcribes English, French, Spanish and Portuguese. In TII’s analysis, Falcon-ASR posts a 20.92% common phrase error charge (WER) throughout 6 Arabic take a look at units. That beats the most effective revealed entry in TII’s leaderboard snapshot by 2.25 factors. For builders constructing Arabic voice merchandise, the catch is entry: at the moment it runs solely as a hosted demo.
TL;DR
- Measurement: 1.6B parameters. Max audio context: Not disclosed (the demo accepts clips as much as 60 seconds and 20 MB).
- Runs on: Not disclosed. No weights or quantized variants are revealed. Entry is a Hugging Face demo, with API and native apps deliberate.
- Efficiency: Lowest Arabic and Emirati WER in TII’s personal comparisons, with stable English outcomes.
- Finest: 20.92% common Arabic WER, 2.25 factors beneath Audar-ASR-V1-Turbo (23.17%).
- Worst: 11.86% WER on Earnings-22, its weakest English take a look at set.
- Backside line (greatest): a compact 1.6B mannequin main bigger Arabic methods in TII’s exams.
- Backside line (worst): no weights, no license, no API but, and all scores are self-reported.
What’s Falcon-ASR?
Falcon-ASR is a speech-to-text mannequin from TII that converts spoken Arabic and 4 different languages into textual content. All 5 languages share one set of weights. The mannequin wants no language flag and outputs textual content within the language spoken. It additionally returns word-level timestamps, linking every phrase to its place within the audio. That is essential for subtitling, searchable recordings and assembly notes.
How does Falcon-ASR deal with Arabic dialects?
Dialectal Arabic has far fewer transcribed sources than Fashionable Customary Arabic (MSA). TII educated Falcon-ASR on Emirati, MSA, different Gulf and Arabic dialects, and English. The target is on a regular basis speech, together with dialect kinds and language switching.
Coaching audio was augmented with background noise, overlapping speech, music, room reverberation and telephony results. Velocity and pitch had been diverse too. The identical therapy was utilized to Emirati recordings.
The mannequin builds on TII’s Falcon3-Audio work. That analysis describes audio-language fashions constructed on instruction-tuned LLMs and Whisper encoders, educated on underneath 30K hours of public audio. Our editorial and technical group at Marktechpost couldn’t discover any revealed Falcon-ASR’s precise structure or coaching knowledge dimension.
How does Falcon-ASR rating on Arabic benchmarks?
TII adopted the protocol of the Open Universal Arabic ASR Leaderboard, run by the ELM Analysis Heart. It ranks methods by equal-weight common WER over 6 take a look at units. TII used the leaderboard’s pinned manifests. Competitor numbers are revealed leaderboard averages checked on 30 September 2026.
| Mannequin | Params | Avg WER (%) | Avg CER (%) |
|---|---|---|---|
| Falcon-ASR | 1.6B | 20.92 | 8.79 |
| Audar-ASR-V1-Turbo | 2.35B | 23.17 | 9.23 |
| Cohere Transcribe Arabic (07-2026) | 2.0B | 25.87 | 11.80 |
| omniASR LLM 7B | 7.0B | 28.32 | 12.52 |
Decrease is best. Notice that Falcon-ASR’s determine is TII’s personal run, not a leaderboard entry. The Audar model card lists its personal common CER as 9.20%.
How does Falcon-ASR carry out on Emirati speech?
TII constructed an inner Emirati and Gulf take a look at set from held-out recordings with human-validated transcripts. It enhances the UAE subset of the general public Casablanca dataset.
| Mannequin | Params | WER (%) | CER (%) |
|---|---|---|---|
| Falcon-ASR | 1.6B | 22.73 | 10.19 |
| Qwen3-Omni-30B-A3B-Instruct | 30B (3B energetic) | 26.80 | 12.72 |
| Audar-ASR-V1-Turbo | 2.35B | 27.89 | 13.75 |
| Cohere Transcribe Arabic | 2.0B | 31.05 | 18.07 |
| Qwen3-ASR-1.7B | 2.0B | 31.52 | 13.35 |
| Audar-ASR-V1-Flash | 0.78B | 32.87 | 15.36 |
Falcon-ASR leads by 4.07 WER factors over Qwen3-Omni. This take a look at set is personal, so it can’t be reproduced. Audar stories 19.4% WER on a distinct Emirati set (Mixat), so the numbers usually are not straight comparable.
How good is Falcon-ASR at English?
On the 7 English take a look at units utilized by the Hugging Face Open ASR Leaderboard, TII stories a 5.74% imply WER. Per-set scores vary from 1.75% on LibriSpeech clear to 11.86% on Earnings-22. For context, Qwen3-ASR-1.7B lists 5.76% imply WER on that leaderboard. French, Spanish and Portuguese are supported, however TII revealed no scores for them.
Comparability: Falcon-ASR vs closest Arabic ASR fashions
| Characteristic | Falcon-ASR | Audar-ASR-V1-Turbo | Cohere Transcribe Arabic | Qwen3-ASR-1.7B |
|---|---|---|---|---|
| Group | TII (UAE) | AudarAI | Cohere | Alibaba Qwen |
| Parameters | 1.6B | 2.35B | 2B | 1.7B (2B on HF card) |
| Structure | Builds on Falcon3-Audio | Audio encoder + LM decoder | FastConformer encoder + Transformer decoder | Constructed on Qwen3-Omni |
| License | Not disclosed | AudarAI Neighborhood License v1.0 | Apache 2.0 | Apache 2.0 |
| Open weights | No | Sure (BF16, GGUF, W4A16) | Sure | Sure |
| Languages | 5 (Arabic incl. Emirati, EN, FR, ES, PT) | 30 (Arabic-first) | Arabic, English | 30 languages + 22 Chinese language dialects |
| Arabic avg WER (6 units) | 20.92 (TII eval) | 23.17 | 25.87 | 33.36 (leaderboard) |
| Arabic avg CER | 8.79 (TII eval) | 9.20 | 11.80 | 12.33 |
| TII Emirati WER | 22.73 | 27.89 | 31.05 | 31.52 |
| English imply WER (Open ASR units) | 5.74 (TII eval) | Not disclosed | Not disclosed | 5.76 |
| Timestamps | Phrase-level | Not disclosed | Not disclosed | By way of Qwen3-ForcedAligner-0.6B |
| Runs on | Not disclosed (hosted demo) | CPU, GPU, edge through GGUF; W4A16 on 12 GB+ GPU | Not disclosed | GPU through Transformers or vLLM |
| Entry | HF demo; API deliberate | Weights + Audar API | Weights + Cohere API | Weights + DashScope API |
Sources: Falcon-ASR figures from TII’s blog. Audar, Cohere and Qwen leaderboard figures from the ELM leaderboard as reproduced on the Audar card. Emirati WER for all fashions is from TII’s inner analysis.
window.addEventListener(‘message’,operate(e){var f=doc.getElementById(‘fasr-frame’);if(f&&e.supply===f.contentWindow&&e.knowledge&&e.knowledge.fasrHeight){f.fashion.peak=e.knowledge.fasrHeight+’px’;}});
Key Takeaways
- Falcon-ASR is a 1.6B Arabic-first speech mannequin from TII with an Emirati focus.
- It stories 20.92% common Arabic WER, 2.25 factors forward of Audar-ASR-V1-Turbo.
- On TII’s Emirati set it beats a 30B Qwen3-Omni by 4.07 WER factors.
- 1 set of weights covers 5 languages, with word-level timestamps.
- No weights, license or API but: take a look at it through the 60-second demo.
Take a look at the Falcon-ASR blog, the Falcon3-Audio paper and check out the demo. Additionally, be happy to observe us on Twitter and don’t neglect to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
The publish TII Releases Falcon-ASR: 1.6B Arabic Speech Model Hits 20.92% WER appeared first on MarkTechPost.

