Introducing Falcon ASR
huggingface ·
The Technology Innovation Institute (TII) has introduced Falcon-ASR, a 1.6 billion parameter speech recognition model designed primarily for Arabic with a specific focus on the Emirati dialect. Developed at TII in Abu Dhabi, the model also supports English, French, Spanish, and Portuguese using a single set of weights without requiring language flags. In evaluations against the Open Universal Arabic ASR Leaderboard, Falcon-ASR achieved an average word error rate of 20.92%, outperforming the previous best published result of 23.17%. Internal testing on Emirati speech yielded a WER of 22.73%, marking the lowest error rate among comparable systems including Qwen3-Omni and Audar-ASR-V1-Turbo. The training data incorporated diverse recording conditions such as background noise, overlapping speech, and telephony effects to enhance real-world applicability. Additionally, the model provides word-level timestamps for transcriptions and demonstrated strong performance on English benchmarks with a mean WER of 5.74%. Users can currently test the model via a Hugging Face Demo Space, with API access and native applications planned for the future.
huggingface