modelscope/FunASR

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

What it solves

FunASR is an industrial-grade speech recognition toolkit designed to provide high-performance, end-to-end ASR (Automatic Speech Recognition) for offline, streaming, and edge deployment. It addresses the need for a flexible toolkit where users can mix and match specialized models for transcription, voice activity detection (VAD), punctuation, and speaker diarization, rather than relying on a single monolithic model.

How it works

The toolkit uses an AutoModel pipeline that coordinates multiple specialized models. For example, a typical production pipeline combines a transcription model (like Paraformer or Fun-ASR-Nano) with a VAD model (fsmn-vad) and a speaker identification model (cam++). It supports various deployment targets, including GPU acceleration via vLLM, OpenAI-compatible API servers, and a C++ runtime based on llama.cpp for high-efficiency CPU and edge deployment using GGUF formats.

Who it’s for

It is intended for developers and engineers building voice-enabled applications, AI agents, and industrial speech services that require low latency, high accuracy (especially for Chinese, English, and Japanese), and the ability to deploy on diverse hardware from high-end GPUs to edge devices.

Highlights

  • High Performance: Claims up to 340x realtime speed with Fun-ASR-Nano and vLLM.
  • Versatile Model Zoo: Includes specialized models for ASR, emotion recognition, audio events, and speaker diarization.
  • Multi-platform Deployment: Supports Python, Docker, and standalone C++ binaries (via llama.cpp) for Windows, Linux, and macOS.
  • Broad Language Support: Offers checkpoints for 31+ languages, with specific strengths in Chinese dialects and regional accents.
  • Agent-Ready: Provides an MCP server for Claude/Cursor and OpenAI-compatible endpoints for LangChain and Dify.

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
    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
    kadirnar/whisper-plus

    A comprehensive audio and video processing toolkit that extends OpenAI's Whisper for transcription, speaker diarization, summarization, and RAG-based video chatting.

  • Project
    modelscope/FunClip

    An open-source automated video clipping tool that uses ASR and LLMs to extract video segments based on text transcripts, speaker identity, or AI prompts.

  • Project
    ggml-org/whisper.cpp

    A high-performance C/C++ port of OpenAI's Whisper model for efficient, offline, on-device automatic speech recognition across multiple platforms.