The Technology Innovation Institute (TII) has introduced Falcon-ASR, a speech-to-text model capable of processing Arabic and four other languages. The model utilizes a single set of weights for all supported languages and does not require language identification flags. It provides word-level timestamps, which are intended for applications such as subtitling and meeting transcription.
TII trained the model on a dataset including Modern Standard Arabic, Emirati, various Gulf dialects, and English. To improve robustness, the training data was augmented with background noise, music, overlapping speech, and telephony effects. The architecture is based on the previous Falcon3-Audio research, which incorporates instruction-tuned large language models and Whisper encoders.
In internal tests conducted by TII, Falcon-ASR achieved a word error rate (WER) of 20.92% on a private Emirati and Gulf test set, outperforming Qwen3-Omni by 4.07 WER points. On the seven English test sets used by the Hugging Face Open ASR Leaderboard, the model recorded a mean WER of 5.74%, with individual results ranging from 1.75% on LibriSpeech clean to 11.86% on Earnings-22.
The model also supports French, Spanish, and Portuguese, although TII has not released performance scores for these languages. The company utilized the Open Universal Arabic ASR Leaderboard protocol for its evaluations, comparing its results against competitor figures published as of September 30, 2026.
