AudarAI/Audar-ASR-V1

Arabic-first generative speech recognition — Audar-ASR-V1 (Flash + Turbo). #1 on the Open Universal Arabic ASR Leaderboard. Model cards, benchmarks & inference.

What it solves

Audar-ASR-V1 provides high-accuracy speech-to-text transcription for Arabic, specifically targeting the challenge of dialectal Arabic (such as Gulf/Emirati, Egyptian, Levantine, and Maghrebi) and code-switched Arabic-English speech. It aims to outperform general-purpose ASR systems on these specific linguistic nuances.

How it works

The project uses a generative approach to speech recognition, treating transcription as audio-conditioned next-token prediction. It is built on an open-weight audio-LLM foundation and adapted using over 300,000 hours of labeled audio. The training process involves a four-stage curriculum ending in KTO preference alignment from native Arabic annotators.

Architecturally, it employs a Whisper-style 128-mel audio encoder paired with a Qwen3 decoder. It is available in two tiers:

  • Flash: A small, real-time model (0.78B parameters) for edge and on-device use.
  • Turbo: A larger, more accurate model (2.35B parameters) for hard dialectal and long-form audio.

Who it’s for

Developers building voice agents, live captioning services, and applications requiring high-precision transcription of Arabic dialects and English, across 30 languages in total.

Highlights

  • State-of-the-Art Performance: Ranked #1 on the Open Universal Arabic ASR Leaderboard for the Turbo tier.
  • Dialect-Faithful: Specifically optimized for major Arabic dialects and code-switching.
  • Flexible Deployment: Supports multiple runtimes including Hugging Face Transformers, GGUF (via llama.cpp), and vLLM for GPU serving.
  • Streaming Capabilities: Implements a "LocalAgreement-2" policy for stable, incremental output with sub-250ms latency.
  • Efficient Scaling: Both tiers share the same prompt interface, allowing developers to switch between Flash and Turbo without changing code.

Related

  • Project
    QwenAudio/Fun-ASR

    A family of high-precision end-to-end speech recognition models optimized for noisy environments, diverse Chinese dialects, and high-throughput production deployment.

  • Project
    salute-developers/GigaAM

    A family of Conformer-based acoustic models for state-of-the-art speech recognition and emotion detection, with strong support for Russian and 70+ other languages.

  • Dispatch
    Fine-Tuning Wav2Vec2-BERT for Low-Resource ASR

    Hugging Face demonstrates how to fine-tune Meta's Wav2Vec2-BERT model for Automatic Speech Recognition (ASR) in low-resource languages, achieving performance comparable to Whisper-large-v3 while being significantly faster and more resource-efficient.

  • Project
    TensorSpeech/TensorFlowASR

    A TensorFlow-based framework for Automatic Speech Recognition that implements various ASR architectures and supports TFLite conversion for efficient deployment.

  • Dispatch
    Fine-Tuning XLS-R for Low-Resource Automatic Speech Recognition

    Hugging Face provides a technical guide on fine-tuning XLS-R, a cross-lingual speech representation model, for low-resource Automatic Speech Recognition (ASR) using the Transformers library.