October 7, 2026
Falcon-ASR transcribes speech without language selection: one set of weights for 5 languages
The model returns text in the language of the recording and timestamps every recognized word.
Falcon-ASR, released on Oct 7, 2026, recognizes 5 languages with a single set of weights. There is no need to specify the language of the recording.
The model has 1.6 billion parameters and converts speech to text. It supports:
- Arabic, including the Emirati dialect. - English. - French. - Spanish. - Portuguese.
Fewer errors. In TII's evaluation, Falcon-ASR made fewer errors than two other models on Arabic speech. WER counts omitted, extra and incorrectly recognized words: a lower percentage means fewer errors.
| TII evaluation | Falcon-ASR, WER | Comparison model | Its WER | | --- | --- | --- | --- | | 6 Arabic test sets, leaderboard snapshot from Sep 30, 2026 | 20.92% | Audar-ASR-V1-Turbo | 23.17% | | Internal Emirati dialect test | 22.73% | Qwen3-Omni-30B-A3B-Instruct | 26.80% |
Across 7 public English test sets, Falcon-ASR's average WER was 5.74%. TII published all these results on Oct 7.
Your own recordings. You can try the model in the official Demo Space on Hugging Face. Word-level timestamps are useful for subtitles and finding a specific point in a recording.
Sources: [Falcon-ASR release and TII evaluation results from Oct 7, 2026](https://huggingface.co/blog/tiiuae/falcon-asr), [official demo for your own recordings](https://huggingface.co/spaces/tiiuae/falcon-asr-demo).
TII plans to offer API access and native Falcon-ASR apps.