🔍 Read the full analysis: Introducing Falcon ASR: A New AI Speech Recognition Model on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
The Technology Innovation Institute in Abu Dhabi has introduced Falcon-ASR, a speech recognition model focused on Arabic and the Emirati dialect. TII reports a 20.92% average word error rate across six Arabic test sets and a 22.73% word error rate in an internal Emirati evaluation; the figures are institute-reported and have not been independently replicated in the source material.
Abu Dhabi’s Technology Innovation Institute (TII) has introduced Falcon-ASR, a 1.6-billion-parameter speech recognition model designed for Arabic, with particular attention to the Emirati dialect, as described in the original analysis. TII reports that the model averaged a 20.92% word error rate across six Arabic test sets and produced its lowest error rates among systems in an internal Emirati comparison; the institute has made a web demo available, while API access and native applications are planned.
On the six test sets in the Open Universal Arabic ASR Leaderboard, TII says Falcon-ASR recorded an equal-weight average word error rate (WER) of 20.92%. The institute compared that result with a 23.17% best published average in the leaderboard snapshot it checked on September 30, 2026. The difference is 2.25 percentage points. Lower WER means fewer word-level transcription errors, but the comparison refers to that dated snapshot rather than a live or complete ranking of all speech-recognition systems.
For Emirati speech, TII reports an internal evaluation result of 22.73% WER and 10.19% character error rate (CER). It says those were the lowest scores among the systems it compared and that the next-best WER, from Qwen3-Omni, was 4.07 percentage points higher. TII describes the evaluation data as held-out Emirati and Gulf recordings with human-validated transcripts. The announcement does not give the evaluation’s size or list every system included. For a related look at compact speech tools, see speech recognition and text-to-speech in under 500 KB.
TII says Falcon-ASR also transcribes English, French, Spanish and Portuguese, using the same model weights for all five languages without requiring users to specify a language. The model provides word-level timestamps, which associate each transcribed word with its position in the audio. People can try it through a Hugging Face demo; TII says API access and native applications are planned but provides no release dates.
Arabic Dialects in Speech Recognition
Speech recognition can be less reliable when people use regional dialects, switch between languages or speak in everyday settings rather than formal broadcasts. Arabic has substantial regional variation, and dialect-specific transcribed material is less available than material for Modern Standard Arabic. Falcon-ASR’s stated focus on Emirati speech, alongside its reported dialect evaluation, addresses a practical challenge for tools intended to transcribe calls, meetings and informal recordings.
If the reported performance carries over to varied real-world audio, the model could help developers build transcription workflows for speakers whose speech is not well represented by formal-Arabic evaluations. Word-level timestamps may also make it easier to locate a phrase within a longer recording. The published results do not establish performance for every dialect, speaker or recording condition, so the announcement is an initial indication rather than proof of broad reliability.
Arabic speech recognition software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How TII Tested the Model
TII says its Arabic benchmark comparison follows the Open Universal Arabic ASR Leaderboard’s protocol, using its pinned manifests and an equal-weight average across six Arabic test sets. The leaderboard is maintained by the ELM Research Center. The institute’s comparison uses published competitor results from a snapshot checked on September 30, 2026, meaning later entries could alter the ranking or average used as a reference. TII has not supplied individual Falcon-ASR scores for each of the six sets in the provided announcement.
The Emirati result comes from a separate internal evaluation, which TII says used held-out Emirati and Gulf recordings with human-validated transcripts. The institute also points to the public Casablanca dataset, which contains a UAE subset. TII says training included Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, as well as English. It reports including noise, overlapping speech, music, room reverberation, telephone effects, and changes in speaking speed and pitch. The model builds on TII’s earlier Falcon3-Audio work.
For English, TII separately reports a 5.74% mean WER across seven public test sets used by the Hugging Face Open ASR Leaderboard. That result concerns a different set of tests and should not be treated as directly comparable with the Arabic or Emirati figures.
““Our aim is to transcribe the words people use in everyday speech, including dialectal forms and switches between languages.””
— Technology Innovation Institute
Emirati dialect transcription tool
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Limits of the Published Results
The scores in the announcement are reported by the model’s developer; the source material does not describe an independent replication. TII has not published a full breakdown of results by Arabic test set, dialect, speaker or recording condition. It also does not specify the size and composition of its internal Emirati evaluation or identify every system included in that comparison. Those details would help readers judge how broadly the figures apply.
The leaderboard comparison is tied to a September 30, 2026 snapshot, and later published results may change the reference point. It is also not clear from the announcement how Falcon-ASR will perform on audio that differs from the evaluation recordings, including speakers with less-represented accents, difficult background noise or different patterns of code-switching. The reported benchmark results alone cannot answer those questions.
multilingual speech recognition API
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Demo Now, Releases Later
Users can test Falcon-ASR through TII’s Hugging Face Demo Space, which the institute says accepts recordings for transcription. That provides a way to inspect outputs on particular audio, although individual demo results do not replace controlled, reproducible evaluations.
TII says it plans to offer API access and native applications, but has not announced dates or further release details. More information on the internal test set, per-dialect results and independent evaluations would help clarify how the reported results translate to practical use. For now, the model’s published scores describe performance on the identified benchmark and internal evaluation material, while broader performance remains to be tested.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Falcon-ASR?
Falcon-ASR is a 1.6-billion-parameter speech recognition model introduced by Abu Dhabi’s Technology Innovation Institute. TII says it supports Arabic, English, French, Spanish and Portuguese, with a particular focus on Arabic and Emirati speech.
What Arabic benchmark result did TII report?
TII reports a 20.92% average word error rate across six Arabic test sets in the Open Universal Arabic ASR Leaderboard protocol. It compared that with a 23.17% best published average in the snapshot it checked on September 30, 2026.
How did Falcon-ASR perform on Emirati speech?
TII reports 22.73% WER and 10.19% CER in an internal evaluation using held-out Emirati and Gulf recordings with human-validated transcripts. The institute says these were the lowest scores among the systems it compared; the source does not provide the evaluation’s size or a full list of systems.
Can people use Falcon-ASR now?
A Hugging Face demo is available for trying the model with recordings. TII says API access and native applications are planned but has not provided a release schedule.
Are the reported results independently verified?
The source material presents the benchmark and internal evaluation figures as TII-reported results and does not describe independent replication. The announcement also leaves out detailed score breakdowns by test set, dialect and recording condition.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
