🔍 Read the full analysis: Discover Falcon ASR: AI For Understanding Spoken Language 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 1.6-billion-parameter speech recognition model focused on Arabic, including Emirati dialect speech. TII reports a 20.92% average word error rate on six Arabic test sets and strong results in an internal Emirati evaluation; the model is available through a Hugging Face demo, while API access and native applications are planned.
The Technology Innovation Institute (TII) in Abu Dhabi has introduced Falcon-ASR, a 1.6-billion-parameter speech recognition model designed for Arabic, with a particular focus on Emirati speech. TII reports a 20.92% average word error rate across six Arabic test sets and says the model scored lowest among the systems in its internal Emirati comparison, as described in the original analysis; users can try it through a Hugging Face demo.
TII says the model transcribes Arabic, English, French, Spanish and Portuguese using the same model weights, without requiring users to specify a language. It returns word-level timestamps, associating recognized words with positions in the audio. According to the institute, training included Emirati Arabic, Modern Standard Arabic, other Gulf and Arabic dialects, and English, as well as recordings affected by noise, overlapping speech, music, reverberation and telephony.
For its Arabic benchmark comparison, TII reports a 20.92% equal-weight average WER across the six test sets used by the Open Universal Arabic ASR Leaderboard. The institute says the best published average in the leaderboard snapshot it checked on 30 September 2026 was 23.17%, a difference of 2.25 percentage points. The leaderboard is maintained by the ELM Research Center; lower word error rate 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 assessment 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. TII also reports a mean 5.74% WER across seven public English test sets used by the Hugging Face Open ASR Leaderboard.
Why Emirati Speech Testing Matters
Speech recognition systems can perform differently on formal, carefully recorded speech than on everyday conversation. Arabic dialects vary across regions, and training and evaluation material is less available for many dialects than for Modern Standard Arabic. A model’s performance on formal broadcasts alone may not predict how well it handles a casual Emirati conversation, a phone call or speakers switching languages.
Falcon-ASR’s reported Emirati evaluation gives developers and potential users a specific data point for a use case that can be missed by broader Arabic benchmarks. If the model performs reliably beyond the test recordings, its timestamps and multilingual transcription could be useful for meetings, calls and longer audio files. The reported scores do not, by themselves, establish performance across all accents, speakers or recording conditions.
Products Worth Considering
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
DJI Osmo Mobile 8 Gimbal with Mic Mini Transmitter, Infinity Black
Apple Watch SE 3 GPS 40mm Smartwatch
As an affiliate, we earn on qualifying purchases.
How TII Tested Falcon-ASR
The Arabic results are based on six test sets combined with an equal-weight average of WER scores. TII compared its result with published leaderboard figures in a snapshot checked on 30 September 2026, rather than claiming a live or permanent ranking. The source material does not give Falcon-ASR’s individual scores for each of the six sets.
TII describes the Emirati results as coming from an internal evaluation with held-out recordings and human-validated transcripts. The institute also points to the public Casablanca dataset, which includes a UAE subset. Falcon-ASR builds on TII’s Falcon3-Audio work. The announcement presents the benchmark and internal-test figures as evidence for the model, but they remain developer-reported results.
““Our aim is to transcribe the words people use in everyday speech, including dialectal forms and switches between languages.””
— Technology Innovation Institute
What the Published Scores Leave Open
The announcement does not provide a full breakdown by Arabic test set, dialect, speaker or recording condition. It also does not specify the size and full composition of the internal Emirati evaluation or list every system included in that comparison. Those details would help readers judge how broadly the reported WER and character error rate apply.
The Arabic leaderboard comparison reflects a snapshot checked on 30 September 2026; later published results could change the relative standing. The material does not describe independent replication of TII’s figures. Real-world performance on recordings that differ from the evaluation data remains unestablished, so the scores should be read as results on the identified tests rather than guarantees for every application.
Demo Available, Releases Pending
Users can currently try Falcon-ASR through TII’s Hugging Face Demo Space, which the institute says accepts recordings for transcription. TII says API access and native applications are planned, but it has not announced release dates. Until those products become available, the demo is the stated route for exploring the model’s output on specific audio.
Further evaluation details, including per-dialect results and independent testing, would help establish how well the benchmark findings transfer to everyday use. For now, the key next developments are the planned access options and any additional evidence TII publishes about performance across speakers and recording conditions.
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.
What Arabic results has TII reported?
TII reports a 20.92% average WER across six Arabic test sets in the Open Universal Arabic ASR Leaderboard. It says its internal Emirati evaluation returned 22.73% WER and 10.19% character error rate.
Can people use Falcon-ASR now?
TII says users can try the model through a Hugging Face demo. API access and native applications are planned, but no release dates were provided.
Are the reported scores independently verified?
The supplied announcement describes the figures as TII-reported results and does not report an independent replication. The internal Emirati evaluation’s size and full system list are also not provided.
Does the benchmark show how Falcon-ASR performs for every Arabic speaker?
No. The reported results apply to the identified test sets and internal recordings. TII has not supplied a complete breakdown by dialect, speaker or recording condition, so performance across all real-world speech remains unclear.
Primary source: Hugging Face · via ThorstenMeyerAI.com
