Falcon ASR : A New Speech Recognition Model Unveiled

Falcon ASR : A New Speech Recognition Model Unveiled
2 min read

The Technology Innovation Institute (TII) in Abu Dhabi has unveiled Falcon-ASR, a 1.6-billion-parameter speech recognition model dedicated primarily to Arabic, with a particular focus on the Emirati dialect. The model also supports English, French, Spanish, and Portuguese.

According to TII, Falcon-ASR achieves an average Word Error Rate (WER) of 20.92% across six Arabic test sets. The best published result in the leaderboard version used by the institute stands at 23.17%. The gap is 2.25 percentage points.

These results are based on the Open Universal Arabic ASR Leaderboard, managed by the ELM Research Center, which ranks systems according to the average WER across six test sets, with equal weight given to each. The leaderboard also indicates the Character Error Rate (CER). For both metrics, the lower the value, the better the result. Competitors’ figures were verified, according to the article, on September 30, 2026.

An Internal Test Focused on Emirati

Public data already includes Emirati, notably the Emirati subset of Casablanca. TII supplemented this coverage with an internal evaluation using additional recordings in Emirati and Gulf dialects, based on tests not used for training and transcripts validated by humans.

On this test, Falcon-ASR achieves a WER of 22.73% and a CER of 10.19%, the lowest values among the compared systems. Its WER is 4.07 points lower than that of Qwen3-Omni, which ranks second.

To cover varied recording conditions, training incorporated background noise, overlapping speech, music, room reverberation, telephony effects, as well as variations in speech speed and pitch. The same processing was applied to Emirati recordings to prepare the model for meetings, calls, and other everyday recordings.

English and Other Languages

Falcon-ASR also transcribes English using the same weights. Across the seven public English test sets used by Hugging Face’s Open ASR Leaderboard, the average WER reaches 5.74%.

French, Spanish, and Portuguese also use the same weights, without needing to specify the language in advance. The output is a transcription in the language actually spoken. The model can also associate each transcribed word with its position in the audio recording, thanks to word-level timestamps.

Falcon-ASR is based on Falcon3-Audio, whose architecture and training method are described in a scientific paper. A demo is available on the project’s Hugging Face space, and TII indicates that API access as well as native applications are planned.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x