Falcon ASR: TII's 1.6B Speech Model for Arabic and Emirati (huggingface.co)

Falcon-ASR, a newly launched 1.6 billion parameter speech recognition model developed by the Technology Innovation Institute (TII) in Abu Dhabi, is specifically tailored for Arabic, with a special emphasis on the Emirati dialect. Achieving an impressive average word error rate (WER) of 20.92% across six Arabic test sets, Falcon-ASR outperformed the best published results by 2.25 percentage points. In internal evaluations focused on the Emirati dialect, it achieved a WER of 22.73%, showcasing unmatched accuracy against competing systems. This model not only captures everyday speech nuances but also provides word-level timestamps linking each transcription to its audio source. Significantly, Falcon-ASR is designed to manage the complexities of spoken Arabic, where dialectal variations pose challenges to traditional models. Its robust training incorporates diverse Arabic dialects, Modern Standard Arabic, and even English, French, Spanish, and Portuguese, all using the same model weights, which facilitates seamless language switching in real-world applications. By addressing common audio challenges, such as background noise and overlapping speech, Falcon-ASR positions itself as a powerful tool for transcribing and understanding multilingual communications in dynamic environments. Future plans for API access and native applications ensure increased accessibility for developers and end-users alike.