Tool Reviews

Meet Falcon ASR

Technology Innovation Institute released Falcon-ASR, a 1.6B-parameter speech recognition model focused on Arabic and the Emirati dialect. It also supports English, French, Spanish, and Portuguese, using shared weights without a language flag. The model reports 20.92% average word error rate on six Arabic leaderboard test sets and 22.73% word error rate, 10.19% character error rate, on an internal Emirati evaluation. It also achieves 5.74% mean word error rate across public English test sets. Falcon-ASR produces word-level timestamps and targets real-world audio with dialects, code-switching, noise, and compression. API access is planned.

Meet Falcon ASR

Key Takeaways

  • TII released Falcon-ASR, a 1.6B-parameter Arabic speech recognition model with a strong focus on the Emirati dialect, while also supporting English, French, Spanish, and Portuguese and producing word-level timestamps.
  • On the Open Universal Arabic ASR Leaderboard, Falcon-ASR achieved an average 20.92% word error rate, 2.25 percentage points better than the best published result in the compared snapshot.
  • In TII’s internal Emirati and Gulf evaluation, the model recorded 22.73% WER and 10.19% CER, outperforming the other compared systems and showing improved accuracy for dialectal Arabic.
  • Falcon-ASR is trained for real-world recording conditions, including background noise, overlapping speech, music, reverberation, and telephony effects, making it suitable for meetings, calls, broadcasts, and everyday audio.
  • The model is available through a Hugging Face demo, with API access and native applications planned, offering a dialect-aware option for developers, researchers, and enterprises.

Meta description: TII releases Falcon-ASR, a 1.6B Arabic speech recognition model with 20.92% WER, Emirati focus, English support and word-level timestamps. API access planned.

Meet Falcon ASR

Tags: Arabic ASR, speech recognition, Falcon-ASR, Emirati dialect, TII

Featured image: An abstract, friendly illustration of waveform shapes, Arabic-inspired geometric patterns and multilingual speech bubbles, with no real people, no logos and no text.

Arabic speech recognition, also called ASR, converts spoken Arabic into text, and the Technology Innovation Institute (TII) in Abu Dhabi has introduced Falcon-ASR, a 1.6 billion parameter model built for that task with a particular focus on the Emirati dialect. The model also supports English, French, Spanish and Portuguese, and it produces word-level timestamps that link each transcribed word to its position in the audio. For developers, researchers and enterprises, the release adds a dialect-aware option for meetings, calls, broadcasts and everyday recordings, where Arabic speakers often move between formal language and regional speech.

Arabic benchmark results

The evaluation follows the protocol used by the Open Universal Arabic ASR Leaderboard, maintained by the ELM Research Center. That leaderboard ranks systems using the equal-weight average word error rate (WER) across six Arabic test sets and also reports character error rate (CER). WER measures how many words are wrong compared with a reference transcript, while CER measures errors at the character level; lower values are better for both.

Using the leaderboard's pinned manifests, Falcon-ASR achieved an average WER of 20.92% across the six Arabic test sets. The best published result in the leaderboard snapshot used for comparison was 23.17%, and TII reports that the new model's average WER is 2.25 percentage points better than that figure. Competitor figures are the published leaderboard averages checked on 30 September 2026.

Public Arabic speech data is not evenly distributed. Modern Standard Arabic (MSA) has more transcribed resources than dialectal Arabic, so models can perform well on news-style speech while struggling with everyday conversation. Falcon-ASR was trained on Emirati, MSA, other Gulf and Arabic dialects, and English, with the aim of transcribing words people actually use, including dialectal forms and changes between languages.

Meet Falcon ASR

Emirati evaluation and dialect focus

Public evaluation data already includes Emirati speech through Casablanca, which has a UAE subset. TII complemented that public coverage with an internal evaluation of additional Emirati and Gulf speech. The internal evaluation used held-out recordings and human-validated transcripts, allowing the team to test transcription accuracy on material that is not limited to the public UAE subset.

On that internal Emirati evaluation, Falcon-ASR recorded 22.73% WER and 10.19% CER. TII reports that it has the lowest WER and CER among the systems compared there. Its WER is 4.07 percentage points below Qwen3-Omni, the next best result. The numbers show improved transcription accuracy at both word and character level on this evaluation.

The Emirati focus is one of the more practical parts of the release. Gulf dialects can differ from MSA in vocabulary, pronunciation and code-switching patterns, and phone recordings add compression, background noise and speaker overlap. A model that performs well only on clean studio audio may fail in those settings. Falcon-ASR is therefore aimed at the messier audio that appears in real deployments.

Multilingual support and timestamps

Falcon-ASR does not limit itself to Arabic. It also transcribes English using the same model weights. In TII's evaluation on the seven public English test sets used by the Hugging Face Open ASR Leaderboard, the model achieved a mean WER of 5.74%. It also supports French, Spanish and Portuguese.

All five languages use the same weights, without requiring a language flag. The output is a transcript in the language spoken. This is useful for recordings where speakers switch between languages or where an application does not know which language will appear in a file.

The model also supports word-level timestamps. Each transcribed word is linked to its position in the audio, which can help users jump to the relevant moment in a recording, build searchable transcripts, create subtitles or review specific utterances in a longer meeting or call.

Meet Falcon ASR

Training conditions and model foundation

TII says Falcon-ASR builds on its Falcon3-Audio work. The architecture and training approach for Falcon3-Audio are described in the paper "Competitive Audio-Language Models with Data-Efficient Single-Stage Training on Public Data".

To prepare the model for different recording conditions, the training data included background noise, overlapping speech, music, room reverberation and telephony effects, along with variations in speed and pitch. TII says the same treatment was applied to Emirati recordings, exposing the model to a range of conditions it may encounter in meetings, calls and other everyday recordings.

The acknowledgments section names the team behind the Falcon-Emirati model for support with Arabic foundation models, pointing to their latest work in the Falcon-Emirati blog post, and credits Mikhail Lubinets for continued support with the compute infrastructure.

Who should use it / who should skip it

A Hugging Face Demo Space lets users try Falcon-ASR and explore its transcription capabilities. API access and native applications are planned. The demo is the starting point for testing the model on recordings that reflect real usage.

For teams that need Arabic transcripts from real recordings, the public benchmark result indicates better performance across six Arabic test sets, while the internal Emirati evaluation points to improved accuracy on a dialect with fewer public evaluation resources. The 5.74% mean WER on the Hugging Face Open ASR Leaderboard English test sets shows the model is not limited to Arabic. The same-weights multilingual design may make deployment easier in mixed-language settings, while planned API and native applications will determine how directly it can be integrated into products.

It may be useful for developers, researchers and enterprises that need dialect-aware Arabic transcription, word-level timestamps, or a single model for Arabic, English, French, Spanish and Portuguese. It may be a poorer fit for teams that require a production-ready API, native application or formal deployment path today, because those features are only described as planned.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.