QwenAudio/Fun-ASR

Fun-ASR speech recognition models, with native Hugging Face Transformers support for Fun-ASR-Nano and separate FunASR, vLLM and llama.cpp deployment paths.

What it solves

Fun-ASR provides high-precision, end-to-end speech recognition (ASR) designed for challenging real-world environments. It specifically addresses the difficulty of transcribing audio with heavy background noise, far-field sound pickup (such as in conference rooms or vehicles), and the complexities of various Chinese dialects and regional accents.

How it works

The project offers a family of 800M parameter models (Fun-ASR-Nano and Fun-ASR-MLT-Nano) that can be deployed in multiple ways:

  • Native Transformers: Integration with the transformers library for standard Python-based inference.
  • High-Throughput Inference: Integration with the vLLM engine for massive batch processing and real-time streaming via WebSockets.
  • Edge Deployment: A self-contained binary via llama.cpp and GGUF formats, allowing for CPU-only inference without a Python runtime.
  • Pipeline Composition: It can be combined with other models like fsmn-vad (Voice Activity Detection) for segmenting long audio and cam++ for speaker diarization (identifying who is speaking).

Who it’s for

Developers and engineers building speech-to-text applications that require high accuracy in noisy environments, support for a wide range of Asian languages and dialects, or the ability to deploy on diverse hardware from high-end GPUs to edge CPUs.

Highlights

  • Extensive Language Support: Supports Chinese (including 7 major dialects and 26 regional accents), English, Japanese, and a multilingual checkpoint covering 31 languages.
  • Noise Robustness: Optimized for far-field and high-noise scenarios, achieving up to 93% recognition accuracy.
  • Specialized Recognition: Includes capabilities for recognizing lyrics in music backgrounds and rap speech.
  • Extreme Performance: vLLM integration enables inference speeds up to 16x faster than native PyTorch.
  • Flexible Deployment: Supports everything from production-grade WebSocket services to standalone GGUF binaries.

Related

  • Project
    modelscope/FunASR

    FunASR is an industrial-grade speech recognition toolkit from Alibaba's ModelScope team. It provides a modular Python API, CLI, and server for offline/streaming ASR, plus VAD, punctuation, speaker diarization, emotion recognition, and edge deployment via llama.cpp. It includes multiple pre-trained models (e.g., Fun-ASR-Nano, SenseVoiceSmall) and supports GPU, CPU, and edge devices.

  • Project
    FireRedTeam/FireRedASR2S

    An industrial-grade all-in-one ASR system that combines speech recognition, voice activity detection, language identification, and punctuation prediction, with specialized support for Chinese dialects.

  • Project
    QwenAudio/SenseVoice

    SenseVoice is a multilingual speech foundation model that provides high-accuracy speech recognition, emotion detection, and audio event identification with low-latency inference.

  • Dispatch
    Qwen3-ASR and Qwen3-ForcedAligner Release

    Qwen has open-sourced Qwen3-ASR (1.7B and 0.6B) and Qwen3-ForcedAligner-0.6B, providing state-of-the-art multilingual speech recognition and non-autoregressive timestamp prediction under the Apache 2.0 license.

  • Project
    istupakov/onnx-asr

    A lightweight Python package for Automatic Speech Recognition using ONNX models, eliminating the need for PyTorch or Transformers for fast deployment on edge and server hardware.