Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
**Audar-ASR-V1** is a family of Arabic-first speech recognition models from AudarAI, and it is worth attention less for raw scale than for the problem it targets. Arabic is one of the hardest languages for ASR: Modern Standard Arabic is what models train on, but almost nobody speaks it conversationally. Real audio is Gulf, Egyptian, Levantine, or Maghrebi dialect, frequently code-switched with English, and general-purpose multilingual models degrade badly on it. Audar-ASR-V1 is built specifically for that gap and covers **30 languages** in total. Architecturally it recasts transcription as **audio-conditioned next-token prediction** over a unified text vocabulary — a language-model decoder rather than the CTC or transducer objectives that dominate production ASR. The stack is a Whisper-style 128-mel audio encoder feeding a Qwen3 decoder with a 30-second context window. The project is explicit that it builds on a permissively-licensed open-weight audio-LLM foundation and that **the contribution is the adaptation, not the foundation**: 300,000+ hours of labeled audio through a four-stage curriculum ending in **KTO preference alignment** from native Arabic annotators. That candor is unusual and makes the claims easier to evaluate. Two tiers ship. **Turbo** (2.35B) ranks **#1 of 36 systems on the Open Universal Arabic ASR Leaderboard at 24.78% average WER**. **Flash** (0.78B) is the strongest small model on that board at **#11 of 36 with 33.31% average WER**, beating systems several times its size. Both share one architecture and one prompt interface, so you can develop against Flash and scale to Turbo without touching code. Note that those WER figures look high in absolute terms — that is the nature of the dialectal Arabic benchmark, not a defect; every system on the leaderboard scores in that range. Deployment is well covered. Flash runs on Transformers, GGUF via llama.cpp, and vLLM; Turbo on GGUF and vLLM. Because vLLM implements the Qwen3-ASR architecture natively, GPU serving needs no custom code. One practical gotcha the README flags: the audio projector must stay BF16 because its `ClippableLinear` layer is numerically sensitive, even though the decoder quantizes normally to Q4_K_M or Q8_0. The significant caveat is licensing. The **repository code is Apache-2.0, but the weights are not open source** — Flash ships under a custom "AudarAI Open v1.0" license and Turbo under "AudarAI Community v1.0." Anyone planning commercial deployment needs to read those terms rather than assume permissive reuse. Adoption is also early: 3 forks and roughly 1,450 combined Hugging Face downloads across both tiers as of publication, with the last repository push on 2026-07-20. This is a credible benchmark leader in an underserved language, not yet a battle-tested community project.
ggml-org
Pure C/C++ port of OpenAI Whisper for edge deployment
CJ Pais
A free, open-source, cross-platform speech-to-text app that transcribes your voice entirely offline — press a shortcut, speak, and have the text pasted into any app.