Transcribe Arabic audio, including regional dialects and Arabic-English code-switching, into text.
Run real-time or on-device transcription for voice agents or live captioning using the smaller Flash model.
Serve the models over an OpenAI-compatible API using vLLM for GPU-based production transcription.
Transcribe long recordings, like meetings, by chunking audio automatically past the model's 30 second context.
| audarai/audar-asr-v1 | nv-tlabs/ardy | bytevisionlab/dreamlite | |
|---|---|---|---|
| Stars | 557 | 561 | 562 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | hard | moderate |
| Complexity | 3/5 | 5/5 | 3/5 |
| Audience | developer | researcher | researcher |
Figures from each repo's GitHub metadata at analysis time.
Model weights download automatically from Hugging Face, running the larger Turbo model well benefits from a GPU, and the GGUF audio projector must stay in BF16 precision.
Audar-ASR-V1 is a pair of speech recognition models built specifically for Arabic, from a company called AudarAI. Rather than the older approach many speech recognition systems use, it works more like a language model, predicting the transcript text token by token based on the audio, and it was trained starting from an existing open-weight audio model, then adapted using over 300,000 hours of labeled audio, mostly Arabic with some English, plus a further training stage that uses feedback from native Arabic annotators. It can transcribe Modern Standard Arabic along with major dialects such as Gulf, Egyptian, Levantine, and Maghrebi Arabic, Arabic mixed with English in the same sentence, English on its own, and 30 languages in total. There are two model sizes. Flash is the smaller model at under a billion parameters, meant for real-time use on regular hardware or at the edge. Turbo is the larger model at a little over two billion parameters, meant for the lowest error rate on harder audio such as long recordings in dialect. On a public leaderboard for Arabic speech recognition covering 36 systems, this repository reports Turbo ranked first by average word and character error rate, and Flash ranked eleventh despite its smaller size. Both models share the same architecture and prompt format, so a project can be built against Flash first and switched to Turbo later without changing code. Model weights are hosted on Hugging Face and download automatically, and the repository includes example scripts for running inference through the Transformers library in Python, through GGUF format with llama.cpp for CPU, GPU, or edge devices, or through vLLM for GPU-based serving with an OpenAI-compatible API. There is also a streaming example that only finalizes a word once two consecutive passes over the audio agree on it, so the transcript text does not change after being shown. The code in this repository is released under the Apache 2.0 license, while the model weights themselves are distributed under AudarAI's own separate license terms, which differ between the Flash and Turbo models and should be checked before commercial use.
Two Arabic-focused speech recognition models, Flash and Turbo, that transcribe Modern Standard Arabic, major dialects, and English.
Mainly Python. The stack also includes Python, Transformers, GGUF.
The repository's code is Apache 2.0, free to use and modify including commercially, the model weights themselves use AudarAI's own separate license terms that differ by model.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.