TII launches Falcon-ASR speech recognition model with focus on Emirati Arabic

Abu Dhabi's Technology Innovation Institute (TII) has released Falcon-ASR, a 1.6-billion-parameter speech recognition model designed to master Arabic and its regional dialects, with a specific focus on the Emirati accent.

The open-source model, launched on the Hugging Face platform, also transcribes English, French, Spanish, and Portuguese using the same set of weights, meaning users do not need to flag which language is being spoken.

Why regional dialects are hard for AI

Recognising spoken Arabic remains a stubborn problem for artificial intelligence. While a standard model can easily transcribe formal news broadcasts in Modern Standard Arabic (MSA), it often stumbles when faced with everyday conversational speech, regional dialects, or low-quality phone recordings. Dialectal Arabic also lacks the vast pools of transcribed training data available for MSA.

To get around this, TII researchers trained Falcon-ASR on a mix of Emirati, MSA, other Gulf dialects, and English. The team deliberately exposed the model to real-world audio challenges, including background noise, overlapping conversations, music, room echoes, and telephone line distortions, alongside variations in speed and pitch.

The model also supports word-level timestamps, allowing users to map every transcribed word to its exact moment in the audio file.

Beating larger models on benchmarks

According to TII's evaluation, the model outperforms several larger rivals. On the Open Universal Arabic ASR Leaderboard, maintained by the ELM Research Center, Falcon-ASR recorded an average Word Error Rate (WER) of 20.92%. This beat the previous top-ranked model in the snapshot they used, Audar-ASR-V1-Turbo, which stood at 23.17% despite having more parameters (2.35 billion).

It also beat Cohere Transcribe Arabic (25.87% WER) and the massive 7-billion-parameter omniASR LLM, which scored 28.32%. For both word and character error rates, lower percentages indicate more accurate transcriptions.

In tests specifically targeting the Emirati dialect, the results were more pronounced. TII tested the model against a mix of public datasets and internal, human-validated recordings. Falcon-ASR recorded a WER of 22.73% on Emirati speech. By comparison, Alibaba's Qwen3-Omni-30B—a model with 30 billion parameters, of which 3 billion are active—recorded a 26.80% error rate.

Multilingual support and availability

Beyond its focus on Arabic, Falcon-ASR achieved a mean WER of 5.74% across seven standard public English test sets, including LibriSpeech and VoxPopuli.

The model builds on the foundations of TII's previous Falcon3-Audio research, which focused on efficient, single-stage training using public data.

TII has launched a public demo space on Hugging Face for users to test Falcon-ASR with their own audio recordings. The institute said it plans to release API access and native applications for the model in the future.