TII Unveils Falcon-ASR, a 1.6B Parameter Arabic Speech Recognition Model

The Technology Innovation Institute (TII) in Abu Dhabi has released Falcon-ASR, a speech-to-text model with 1.6 billion parameters. This release introduces a system specifically optimized for Arabic speech, with a particular emphasis on the Emirati dialect, according to a report by MarkTechPost.

The model supports transcription for Arabic, English, French, Spanish, and Portuguese using a single set of weights. Falcon-ASR does not require a language flag to function, automatically detecting the spoken language and returning text in the corresponding format. It also generates word-level timestamps, which link specific words to their position in the audio file.

Technical Performance and Access Constraints

In TII’s internal evaluation, Falcon-ASR recorded a 20.92% average word error rate (WER) across six Arabic test sets. This figure represents a 2.25-point improvement over Audar-ASR-V1-Turbo, which posted a 23.17% WER in the same comparison snapshot. However, the model’s performance varied by language.

On English benchmarks, the model achieved an 11.86% WER on the Earnings-22 test set, which TII identified as its weakest English result. Despite these mixed results in English, the model led larger Arabic systems in TII’s own tests. All reported scores are self-reported by the institute, as independent validation has not been detailed.

Architecture and Training Methodology

Falcon-ASR builds upon TII’s previous Falcon3-Audio research, which described audio-language models constructed on instruction-tuned large language models and Whisper encoders. The MarkTechPost report notes that the exact architecture and total training data size for Falcon-ASR have not been published.

The training process included audio augmentation techniques to simulate real-world conditions. Developers applied background noise, overlapping speech, music, room reverberation, and telephony effects to the training data. Speed and pitch variations were also introduced to enhance robustness. The model was trained on a mix of Emirati, Modern Standard Arabic, other Gulf dialects, and English recordings.

Dialectal Arabic typically has fewer transcribed resources available than Modern Standard Arabic. TII designed Falcon-ASR to handle everyday speech patterns, including dialect forms and code-switching between languages. The objective was to improve transcription accuracy for under-resourced dialectal varieties.

Availability and Deployment Limitations

Currently, Falcon-ASR is accessible only through a hosted demo on Hugging Face. The demo accepts audio clips up to 60 seconds in length and a maximum file size of 20 MB. The maximum audio context supported by the model has not been disclosed.

TII has not released the model weights, quantized variants, or a software development kit (API). Native applications and API access are planned for future release but have no confirmed timeline. The absence of a published license or weights means developers cannot yet deploy the model locally or integrate it into proprietary products.

What should you do?

Security and technology professionals monitoring speech recognition developments should track the Hugging Face demo for performance benchmarks. Since no weights are available, teams cannot assess the model’s security or data handling practices independently. Organizations relying on Arabic transcription should evaluate the hosted demo for accuracy in their specific dialects before committing resources. Watch for official announcements regarding API release and licensing terms.

Background: Large language models

Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.

Read the full guide: Large Language Models vs Generative AI: The Distinction and Security Implications

What to do and how to stay safe: TII

• Monitor the Hugging Face demo to evaluate transcription accuracy for your specific use case and dialect.
• Await the release of model weights and a license to assess security implications and deployment feasibility.
• Track official communications from TII for updates on API availability and native application support.
• Verify self-reported performance metrics against independent benchmarks once public testing becomes possible.

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Can I download the Falcon-ASR model weights for local deployment?

No, TII has not released the model weights or quantized variants, so local deployment is not currently possible.

Which languages does Falcon-ASR support?

The model supports Arabic, English, French, Spanish, and Portuguese using a single set of weights.

How does Falcon-ASR compare to other Arabic models?

It reports a 20.92% average word error rate on Arabic tests, beating the best published entry in TII’s snapshot by 2.25 points.