Falcon ASR Release

Falcon-ASR is a 1.6 billion parameter automatic speech recognition (ASR) model developed by the Technology Innovation Institute (TII) in Abu Dhabi. It is specifically optimized for Arabic, with a high degree of focus on the Emirati dialect, while also providing support for English, French, Spanish, and Portuguese.

State-of-the-Art Arabic and Emirati Performance

Falcon-ASR outperforms existing systems on both public and internal Arabic benchmarks. In evaluations using the Open Universal Arabic ASR Leaderboard protocol across six test sets, Falcon-ASR achieved an average Word Error Rate (WER) of 20.92%, improving upon the previous best published result of 23.17% by 2.25 percentage points.

For the Emirati dialect, TII conducted internal evaluations using held-out recordings and human-validated transcripts to supplement public data from the Casablanca dataset. The results demonstrate that Falcon-ASR achieves the lowest WER and Character Error Rate (CER) among compared systems:

  • Emirati WER: 22.73%
  • Emirati CER: 10.19%

In this internal evaluation, Falcon-ASR's WER was 4.07 percentage points lower than the next best performing model, Qwen3-Omni.

Multilingual Capabilities and English Benchmarks

Falcon-ASR uses a single set of model weights to transcribe five different languages without requiring a language flag. In addition to Arabic, the model supports:

  • English
  • French
  • Spanish
  • Portuguese

When evaluated on seven public English test sets from the Hugging Face Open ASR Leaderboard, Falcon-ASR recorded a mean WER of 5.74%.

Robustness to Real-World Audio Conditions

To ensure the model functions effectively in everyday settings such as meetings and phone calls, TII trained Falcon-ASR on audio containing various environmental distortions. The training data included:

  • Background noise and overlapping speech
  • Music and room reverberation
  • Telephony effects
  • Variations in speech speed and pitch

Technical Foundation and Features

Falcon-ASR is built upon the Falcon3-Audio framework. Its architecture and training methodology are detailed in the research paper Competitive Audio-Language Models with Data-Efficient Single-Stage Training on Public Data.

Key functional features include:

  • Word-level timestamps: The model links each transcribed word to its specific position in the audio file.
  • Language Agnosticism: The model automatically detects and transcribes the spoken language from the supported set using the same weights.

Availability

Falcon-ASR is currently available for testing via a Hugging Face Demo Space. TII has announced that API access and native applications are planned for future release.

Sources