🔍 Read the full analysis: How Does Falcon ASR Fit Into The World Of AI? 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
Abu Dhabi’s Technology Innovation Institute has introduced Falcon-ASR, a 1.6-billion-parameter speech recognition model designed for Arabic, particularly Emirati dialect, and five supported languages. TII reports benchmark results on Arabic and Emirati evaluations, but the Emirati figures are from an internal test and independent replication is not described.
The Technology Innovation Institute (TII) in Abu Dhabi has introduced Falcon-ASR, a 1.6-billion-parameter speech recognition model built with a focus on Arabic and the Emirati dialect, as detailed in the original analysis. TII reports a 20.92% average word error rate across six Arabic test sets and says Falcon-ASR had the lowest error rates among the systems in its internal Emirati comparison; the results could interest developers working on transcription for Arabic speech, but the Emirati evaluation has not been independently replicated in the material provided.
TII says Falcon-ASR can transcribe Arabic, English, French, Spanish and Portuguese using the same model weights, without requiring users to specify a language. The institute says its training included Emirati, Modern Standard Arabic, other Gulf and Arabic dialects, and English. It also included audio conditions such as background noise, overlapping speech, music, reverberation and telephone effects, according to TII.
For Arabic, TII reports an average word error rate (WER) of 20.92% across six test sets in the Open Universal Arabic ASR Leaderboard. In the leaderboard snapshot checked by TII on September 30, 2026, the best published average was 23.17%, a 2.25 percentage-point difference. The leaderboard, maintained by ELM Research Center, averages the six sets with equal weight. Lower WER means fewer word-level transcription errors. TII says it followed the leaderboard protocol and used its pinned manifests.
For Emirati speech, TII reports 22.73% WER and 10.19% character error rate in an internal evaluation using held-out Emirati and Gulf recordings with human-validated transcripts. It says these were the lowest scores among the systems it compared, and that the next-best WER, from Qwen3-Omni, was 4.07 percentage points higher. Falcon-ASR also returns word-level timestamps, which associate each transcribed word with its position in the audio.
Why Emirati Speech Results Matter
Speech recognition tools can perform differently across accents, dialects and recording environments. Arabic is spoken in many regional forms, while training and evaluation material is less available for some dialects than for Modern Standard Arabic. That can make performance on formal speech a poor guide to how a system handles everyday conversation, calls or meetings.
Falcon-ASR’s stated emphasis on Emirati and Gulf speech addresses a practical need for people building transcription services in the region. If its reported performance carries over to varied real-world recordings, the model could help with tasks such as searching meeting transcripts or locating speech in longer audio. Word-level timestamps may make that work easier by pointing users to specific moments in a recording.
The results are still bounded by the evaluations TII describes. The six-set Arabic average and internal Emirati test provide evidence about those test materials, not a guarantee of accuracy for every speaker, accent, language-switching pattern or recording condition. Developers and users will need real-world testing before treating the scores as a dependable measure for a particular application.
Arabic speech recognition software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How TII Benchmarked Falcon-ASR
The Arabic result is a comparison against a dated leaderboard snapshot, not a live ranking. TII checked the published results on September 30, 2026, and reports that the six test sets were averaged equally. The source material does not include Falcon-ASR’s individual score on each set, making it difficult to see whether its average reflects consistently similar performance or stronger results on particular tests.
TII describes the Emirati assessment separately as an internal evaluation using held-out recordings and human-validated transcripts. It also points to the public Casablanca dataset, which has a UAE subset. The announcement does not say that the internal evaluation is equivalent to a public benchmark or provide enough detail to directly compare the two.
The institute says Falcon-ASR builds on its Falcon3-Audio work. Separately, it reports a mean WER of 5.74% on seven public English test sets used by the Hugging Face Open ASR Leaderboard. TII has made a Hugging Face demo available for users to try with recordings. API access and native applications are described as planned, without release dates.
“Our aim is to transcribe the words people use in everyday speech, including dialectal forms and switches between languages.”
— Technology Innovation Institute
As an affiliate, we earn on qualifying purchases.
Limits of the Published Evaluations
The announcement does not provide a full breakdown of results by Arabic test set, dialect, speaker or recording condition. It also does not specify the size and composition of the internal Emirati evaluation or list every system included in that comparison. Those details would help readers judge how broadly the reported scores apply.
The Arabic comparison reflects leaderboard results available on September 30, 2026; later submissions could change the relative standing. The figures are reported by TII, and the source material does not describe an independent replication. Performance outside the identified test recordings, including for speakers and conditions not represented in them, remains uncertain.
As an affiliate, we earn on qualifying purchases.
Demo Access and Planned Releases
People can currently try Falcon-ASR through TII’s Hugging Face Demo Space, which the institute says accepts users’ recordings for transcription. TII has also said API access and native applications are planned, but it has not announced when they will be available.
Further information that would help assess the model includes per-test and per-dialect results, more detail about the Emirati evaluation, and independent testing. Until those details or broader product releases are available, the demo offers a way to examine the system on particular recordings, while the published benchmark numbers should be understood as results from the evaluations TII identifies.
Emirati dialect transcription tool
As an affiliate, we earn on qualifying purchases.
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.
How did Falcon-ASR score on Arabic tests?
TII reports a 20.92% average word error rate across six Arabic test sets in a leaderboard snapshot checked on September 30, 2026. The reported average was 2.25 percentage points below the best published average in that snapshot.
What are the reported results for Emirati speech?
TII reports 22.73% WER and 10.19% character error rate in an internal evaluation using held-out Emirati and Gulf recordings with human-validated transcripts. The announcement does not describe an independent replication.
Can people try Falcon-ASR now?
Yes. TII says a Hugging Face demo is available for trying the model with recordings. API access and native applications are planned, but no release dates were given.
Do the reported scores show how Falcon-ASR will perform for every speaker?
No. The published figures describe specified benchmark and internal evaluation materials. TII has not provided a complete breakdown by dialect, speaker or recording condition, so performance across other real-world audio remains uncertain.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
