TII Releases Falcon-ASR: 1.6B Parameter Model for Arabic and English Speech Recognition

The Technology Innovation Institute (TII) has released Falcon-ASR, a 1.6 billion parameter speech recognition model optimized for Arabic, with a specific focus on the Emirati dialect. The model also supports English, French, Spanish, and Portuguese using a single set of weights without requiring language flags.

Benchmark Performance

In evaluations against the Open Universal Arabic ASR Leaderboard, Falcon-ASR achieved an average Word Error Rate (WER) of 20.92% across six Arabic test sets. This is 2.25 percentage points better than the best previously published result of 23.17% (Audar-ASR-V1-Turbo). The model also recorded a Character Error Rate (CER) of 8.79%.

On TII's internal evaluation for Emirati speech, Falcon-ASR recorded the lowest error rates among compared systems:

• WER: 22.73% (4.07 percentage points lower than Qwen3-Omni)
• CER: 10.19%

For English, the model achieved a mean WER of 5.74% on the seven public test sets used by the Hugging Face Open ASR Leaderboard, including 1.75% on LibriSpeech clean.

Technical Details and Availability

Falcon-ASR builds on the Falcon3-Audio architecture and was trained on Emirati, Modern Standard Arabic (MSA), other Gulf and Arabic dialects, and English. The training data included variations in background noise, overlapping speech, music, reverberation, and telephony effects to improve robustness in real-world conditions. The model supports word-level timestamps for transcriptions.

A demo is available on Hugging Face, with API access and native applications planned for future release.