OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Audar Asr V1

Audar-ASR-V1 is a family of generative speech-recognition models that converts audio into text, with a focus on Arabic and its dialects. The repository documents Flash and Turbo variants, model cards, benchmarks and inference routes using Transformers, llama.cpp and vLLM.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Audar Asr V1 — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use the repository when you need an Arabic-focused ASR model and want documented local inference or serving options. Flash is presented for real-time and edge scenarios; Turbo is presented for accuracy-focused use on difficult dialectal and long-form audio.

Who it is for

  • Developers building Arabic speech-to-text workflows
  • Teams comparing Arabic ASR model variants
  • Practitioners working with dialectal or Arabic–English code-switched recordings

Documented uses

  • Transcribe short Arabic or English audio clips
  • Process longer recordings using the documented chunking script
  • Explore incremental transcription with the repository's streaming example
  • Compare Flash and Turbo using the repository's benchmark materials
Before you choose

Requirements & limits

The repository code is identified as Apache-2.0 licensed. Flash weights use AudarAI Open v1.0; Turbo weights use AudarAI Community v1.0. Review the relevant model license separately from the code license.

  • Flash has 0.78 billion parameters and is described for real-time and edge use; Turbo has 2.35 billion parameters and is positioned for accuracy.
  • Transformers is documented for Flash, while llama.cpp and vLLM are documented for both tiers.
  • The Python reference helper handles clips up to 30 seconds; a separate script chunks long-form audio.
  • The llama.cpp route requires Qwen3-ASR support and a BF16 audio projector.
  • Benchmark results are reported in the repository and may not predict performance on a specific recording.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Transcribe a short clip with Transformers and Flash

Use the documented Python example to transcribe a short audio file with the Flash model.

Before you start

  • Python environment with dependencies from examples/requirements.txt
  • An audio clip no longer than 30 seconds
  • Access to download the Flash weights from Hugging Face

Steps

  1. Install the dependencies with pip install -r examples/requirements.txt.

  2. Run python examples/transcribe.py clip.wav for Arabic with automatic dialect handling.

  3. For an English recording, run python examples/transcribe.py english.wav --lang en.

Expected outcome

The example transcribes the supplied audio clip to text.

Guide limitations

  • The documented reference helper accepts clips up to 30 seconds.
  • This recipe covers the repository's Flash Transformers route.
Documentation 1 ↗
Source-based practical guide

Run GGUF inference with llama.cpp

Prepare the documented llama.cpp route for inference with a Flash or Turbo GGUF model.

Before you start

  • A recent llama.cpp build with Qwen3-ASR support
  • The selected model's GGUF decoder and audio projector files
  • The audio projector in BF16 format

Steps

  1. Build a recent llama.cpp version with Qwen3-ASR support.

  2. Use the repository's GGUF inference helper with an audio file and the selected tier, such as ./examples/gguf_infer.sh clip.wav turbo.

  3. Keep the audio projector BF16 when preparing the model files.

Expected outcome

The documented llama.cpp inference path is set up to accept an audio clip and selected model tier.

Guide limitations

  • The repository warns that the audio projector's ClippableLinear is numerically sensitive and must remain BF16.
  • This guide describes the documented inference setup, not a verified run.
Documentation 1 ↗
COMPARE OPTIONS

Compare options for Model inference and serving

These projects document a common task and share a project role. This is a source-based comparison, not an execution benchmark or a guarantee of interchangeability.

ProjectDocumented purposePlatformAccess and requirements
Audar Asr V1Transcribe short Arabic or English audio clipsLinuxThe repository code is identified as Apache-2.0 licensed. Flash weights use AudarAI Open v1.0; Turbo weights use AudarAI Community v1.0. Review the relevant model license separately from the code license.
Pulse Of MotionEstimate a video's physical frame rate from its visual dynamics.LinuxThe repository is public. Its README describes a released checkpoint that downloads from Hugging Face on first inference, with a manual download option. The strongest internal model is not fully released; the README directs interested users to contact the project about access. No license information is provided in the repository evidence.
Neural Image AssessmentTrain the model using AVA images and rating annotations.Cloud, Linux, macOS, Windows
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
14 October 2026
Original website title
AudarAI/Audar-ASR-V1 · Official homepage title · retained in its source language
How we review sourcesSuggest a correction