The Technology Innovation Institute (TII) in Abu Dhabi has released Falcon-ASR, a speech recognition model with 1.6 billion parameters that is designed mainly for Arabic, with a particular emphasis on the Emirati dialect. According to a blog post published on Hugging Face by TII, the model also transcribes English, French, Spanish and Portuguese, and it can attach timestamps to each transcribed word. The institute says Falcon-ASR outperforms the best published result on a widely used Arabic benchmark and posts the lowest error rates among the systems it compared on an internal Emirati test.
Why spoken Arabic is difficult to transcribe
Arabic is not a single spoken form. Modern Standard Arabic, the formal variety used in news broadcasts, official speeches and most written material, sits alongside a wide range of regional dialects that differ in vocabulary, pronunciation and grammar. A system trained mostly on formal broadcast speech can perform well on a newsreader and then falter when a speaker switches to a local dialect, or when the audio comes from a noisy street or a phone call.
The problem is partly a data problem. Dialectal Arabic has far fewer transcribed recordings than Modern Standard Arabic, which makes it harder to train models and to measure their accuracy. TII says it trained Falcon-ASR on Emirati speech, Modern Standard Arabic, other Gulf and Arabic dialects, and English. The stated goal is to capture the words people actually use in everyday conversation, including dialect forms and speakers who move between Arabic and English within a single sentence.
What the benchmark results show
The main headline figure concerns the Open Universal Arabic ASR Leaderboard, which is maintained by the ELM Research Center. The leaderboard ranks systems by their average word error rate (WER) across six test sets, giving each set equal weight, and it also reports character error rate (CER). In both measures, a lower number means better performance.
TII reports that Falcon-ASR reached an average WER of 20.92 percent across the six Arabic test sets. The best published result in the leaderboard snapshot the institute used was 23.17 percent, so the gap is 2.25 percentage points. The competitor figures were checked against the leaderboard on 30 September 2026, and TII says it ran its own evaluation on the same six benchmarks using the leaderboard's pinned manifests, which is intended to keep the comparison consistent.
Results on the Emirati evaluation
Public Arabic datasets already include some Emirati speech. The Casablanca collection, for example, has a subset recorded in the UAE. TII says it added an internal evaluation of further Emirati and Gulf speech, using held-out recordings and transcripts checked by human reviewers. On this test, Falcon-ASR achieved a WER of 22.73 percent and a CER of 10.19 percent.
According to TII, Falcon-ASR recorded the lowest word and character error rates among the systems in its comparison. Its WER was 4.07 percentage points lower than that of Qwen3-Omni, which ranked second. It is worth noting that this internal test was designed and run by TII itself, so its results have not yet been verified by an independent party. Readers should weigh it accordingly, although the public leaderboard comparison offers a more standardised point of reference.
Training for real-world recording conditions
Speech models often perform well in a studio and poorly in the field. To address this, TII says it trained Falcon-ASR on audio with background noise, overlapping voices, music, room echo and the compression effects typical of telephone lines. It also varied the speed and pitch of recordings. The same treatment was applied to the Emirati recordings, so the model should be better prepared for meetings, phone calls and casual everyday audio.
The model is also designed to handle languages mixed together in one conversation. This matters in the Gulf, where speakers often combine Arabic and English in the same exchange, particularly in business and technical settings.
English, French, Spanish and Portuguese
Falcon-ASR transcribes English with the same model weights used for Arabic. On seven public English test sets used by the Hugging Face Open ASR Leaderboard, TII reports a mean WER of 5.74 percent. The model also supports French, Spanish and Portuguese. All five languages share one set of weights and do not require the user to specify the language in advance. The output is a transcript in whichever language is spoken.
For developers, a single multilingual model can simplify deployment compared with maintaining separate systems for each language. Word-level timestamps add practical value for subtitling, searching recordings and reviewing meetings, since each transcribed word can be tied to its exact position in the audio.
Availability and what comes next
TII has put a demonstration version of the model on a Hugging Face Space, which lets people test transcription with their own recordings. The institute says API access and native applications are planned but has not given dates. Falcon-ASR builds on the earlier Falcon3-Audio work, whose architecture and training approach are described in a separate research paper.
The release adds to Abu Dhabi's growing body of work on Arabic language models. Whether its advantage holds up under independent testing, particularly on the Emirati evaluation, will be the key question for researchers and companies considering it for call centres, media archives and government services.
Related: Pulse Local Keeps Your Budget Chat on Your Own Machine · LFM2.5-24B-A2B: Liquid AI's Open Reasoning Model for On-Device Agents
Comments (0)