TII unveils Falcon-ASR: 1.6B arabic speech model achieves 20.92% word error rate

The Technology Innovation Institute (TII) in Abu Dhabi has launched Falcon-ASR, a 1.6B parameter Arabic speech recognition model that also supports English, French, Spanish, and Portuguese. The model achieves a 20.92% average word error rate (WER) across six Arabic test sets, surpassing the previous best entry on TII’s leaderboard by 2.25 points. Falcon-ASR operates without a language flag and generates text in the spoken language, while also providing word-level timestamps for applications like subtitling and meeting notes. TII trained Falcon-ASR on Emirati, Modern Standard Arabic (MSA), and other Gulf dialects, along with English, using augmented audio data that includes background noise, overlapping speech, and room reverberation. The model is based on TII’s Falcon3-Audio research, which leverages instruction-tuned large language models and Whisper encoders. TII’s internal evaluation also includes a private Emirati test set, complementing the UAE subset of the Casablanca dataset. On the Hugging Face Open ASR Leaderboard, Falcon-ASR reports a 5.74% mean WER across seven English test sets, with scores ranging from 1.75% on LibriSpeech to 11.86% on Earnings-22. While TII did not publish scores for French, Spanish, and Portuguese, the model’s performance on Arabic and English highlights its potential for voice-driven applications.