Introducing Falcon ASR
Falcon‑ASR is a 1.6 B‑parameter multilingual speech‑recognition model that achieves a 20.92 % average Arabic WER—2.25 pp better than the prior leaderboard best—and records a 22.73 % WER (10.19 % CER) on internal Emirati tests, the lowest among compared systems. It also delivers a 5.74 % mean WER on seven public English test sets and supports French, Spanish, and Portuguese with a single weight set and word‑level timestamps. Engineers can leverage the model for dialect‑rich Arabic and multilingual transcription without language‑specific flags, improving accuracy in noisy, telephony, and conversational scenarios.
⚡ Key Takeaways
- Arabic average WER = 20.92 % (best published = 23.17 %); Emirati internal WER = 22.73 % (4.07 pp below Qwen3‑Omni)
- Single‑weight multilingual architecture covering Arabic (including Emirati, MSA, Gulf dialects), English, French, Spanish, Portuguese
- Trained with background noise, overlapping speech, music, reverberation, telephony effects, speed and pitch variations
- Provides word‑level timestamps linking each word to its audio position
- No language flag required; model outputs transcript in the spoken language
The model’s superior dialectal Arabic performance and multilingual capability let production systems replace separate language‑specific ASR pipelines, reducing maintenance overhead while delivering lower error rates in real‑world noisy audio.
✅ Practical Steps
- Integrate Falcon‑ASR weights into your inference pipeline to obtain transcriptions with word‑level timestamps.
- Feed audio recordings (including noisy or telephony conditions) directly to the model without setting a language flag; the output will be in the spoken language.
- Evaluate the model on your target dialects using WER/CER metrics to confirm the reported improvements.
Want the full story? Read the original article.
Read on Hugging Face Blog ↗