Introducing Falcon ASR

What changed

Arabic WER: 20.92% · Parameters: 1.6B · Emirati WER (TII evaluation): 22.73% We’re introducing Falcon-ASR, our 1.6 billion parameter speech recognition model for Arabic, with a particular focus on the Emirati dialect. Developed at the Technology Innovation Institute (TII) in Abu Dhabi, it also supports English, French, Spanish and Portuguese. In our evaluation, Falcon-ASR achieved an average word error rate of 20.92% across six Arabic test sets, compared with the best published result of 23.17% in the leaderboard snapshot we used.

Why it matters

A concrete addition to Practical AI: Tools, Models & Frameworks: it changes what's available to builders today rather than being general commentary.

How it compares

Related prior coverage to compare against:

• Introducing the FFASR Leaderboard: Benchmarking ASR in the Real World
• Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers
• Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance

• Introducing the FFASR Leaderboard: Benchmarking ASR in the Real World similarity 0.31
• Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers similarity 0.28
• Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance similarity 0.27