Audar-ASR-V1
A family of Arabic-first, generative speech-recognition models that treat transcription as audio-conditioned next-token prediction rather than CTC decoding, covering Modern Standard Arabic, major dialects, code-switched Arabic-English and English.
Audar-ASR-V1 is built on an openly licensed audio-LLM foundation and adapted through more than 300,000 hours of labeled audio, mostly Arabic with some English, using a four-stage curriculum that ends in preference alignment from native Arabic annotators. It transcribes MSA, dialects such as Gulf, Egyptian, Levantine and Maghrebi Arabic, code-switched Arabic-English, and English, thirty languages in total.
It ships in two tiers: a 0.78-billion-parameter Flash model for real-time and on-device use, and a 2.35-billion-parameter Turbo model tuned for accuracy on hard dialectal and long-form audio. Both share one Whisper-style audio encoder over a Qwen3 decoder with a 30-second context, so code written against Flash also runs against Turbo.
Both models can be run through Hugging Face Transformers, GGUF with llama.cpp, or vLLM, with weights downloaded automatically from Hugging Face.