TII Releases Falcon-ASR Arabic Speech Recognition Model

What changed

The Technology Innovation Institute (TII) in Abu Dhabi has released Falcon-ASR, a speech recognition model with 1.6 billion parameters that converts spoken language to text. The model prioritizes the Emirati dialect while also handling Modern Standard Arabic, English, French, Spanish and Portuguese through a single set of weights. Word-level timestamps are included in the output, linking each word to its position in the audio.

According to TII’s evaluation across 6 Arabic test sets following the Open Universal Arabic ASR Leaderboard protocol, Falcon-ASR achieved a 20.92% average word error rate (WER). This result exceeds the best previously published entry on that leaderboard by 2.25 points. On TII’s internal Emirati test set built from held-out recordings with human-validated transcripts, the model outperforms a 30-billion-parameter Qwen3-Omni system by 4.07 WER points. For English, the company reports a mean 5.74% WER across 7 test sets from the Hugging Face Open ASR Leaderboard, ranging from 1.75% on LibriSpeech clean to 11.86% on Earnings-22.

Access remains limited for developers. Model weights are not published. No license has been disclosed. The only available interface is a hosted demo on Hugging Face that accepts audio clips up to 60 seconds and 20 MB in size. TII states that an API and native applications are planned but has provided no timeline for their release.

Why it matters

Arabic speech recognition has historically received less research attention and model availability compared to English. Most published systems focus on Modern Standard Arabic or broader language families. Emirati and other Gulf dialects have limited transcribed resources, making specialized models rare. Falcon-ASR addresses this gap by offering a system trained specifically on Emirati speech patterns.

For teams building voice products serving the UAE or broader Emirati-speaking audiences, a compact 1.6B model tuned for this dialect could reduce transcription errors compared to general-purpose systems. The multilingual capability also eliminates the need to run separate models for language switching, which is common in mixed-dialect environments.

The model’s practical value depends on three significant constraints. First, without published weights, enterprises cannot run inference on their own infrastructure or integrate the system into private applications. This limits adoption to organizations willing to rely on third-party hosting. Second, no API means developers currently cannot call the model programmatically. Transcription must occur through the demo interface, which is unsuitable for production workflows. Third, all benchmark scores are TII’s own runs on internal and public test sets. Independent verification is not yet possible.

Teams evaluating Falcon-ASR face a timing question: should they test the demo now to validate performance for their use case, or wait for weights and API availability? The answer depends on project urgency and risk tolerance for unverified benchmarks.

What to test

Before considering Falcon-ASR for production, examine the following:

• Run dialect-specific samples through the 60-second demo. Does Emirati performance translate to actual speech patterns, noise levels and speaker compositions in your environment?
• Test code-switching and mixed-language utterances, especially Arabic-English pairs common in Gulf business contexts.
• Compare error patterns on domain-specific vocabulary such as financial, medical and technical terms against the published benchmarks, which use general speech.
• Confirm latency and reliability on the hosted demo with realistic audio quality, not just studio recordings.
• Request a timeline from TII for weights release, API availability and the licensing model. Vendor timelines frequently shift.
• Ask TII for third-party evaluation data or documentation on how to reproduce their benchmark numbers independently.
• Examine TII’s training data composition. The company states training included under 30,000 hours of public audio but has not published exact dataset sizes or sources. This is limited information for a production-grade model.
• Inquire about TII’s versioning strategy for Falcon-ASR. Will the model receive updates? How will performance regression be managed?

The conclusion

Falcon-ASR fills a real gap in Arabic and Emirati speech recognition. The reported 20.92% WER on Arabic test sets and 4.07-point Emirati advantage over a much larger competitor suggest the specialization delivers results. A 1.6B model is small enough to run on modest hardware if weights become available.

The current access model represents a significant barrier. A 60-second demo validates the concept but does not enable adoption. Teams building voice products need either self-hosted weights or a production API. Neither exists today.

TII’s disclosure of benchmark numbers without independent verification is standard practice for new models but creates uncertainty. Leaderboard entries are reproducible; internal test sets are not. These scores should be treated as encouraging but not definitive until third parties report results.

Watch for three developments: announcement of a weights release date with licensing terms, launch of a public API, and publication of training data details. Each removes a barrier to real-world deployment. Until then, Falcon-ASR remains a promising proof of concept worth testing in the demo but not yet ready for production voice pipelines.

Related reading

• Mistral Large 4: Europe’s 1 Trillion-Parameter Model Launches
• vLLM Decision 3.0: Five Open Decision Models, 0.8B to 27B
• OrcaCyber Zero 1.5 Cybersecurity Model Launches With 1M Context