Verified project record

modelscope/FunASR

FunASR is an industrial-grade speech recognition toolkit supporting offline, streaming, and edge deployment for ASR, VAD, punctuation, speaker diarization, and emotion recognition. It provides a multi-model alternative to Whisper and cloud APIs, offering up to 340x realtime transcription speeds across 52 languages while running on CPU and edge devices.

Project overview

The toolkit achieves up to 340x realtime factors for speech recognition, runs faster on CPU than Whisper runs on GPU, and bundles speaker identification, emotion recognition, and VAD without external dependencies.

Project type: MCP · Model Runtime · Audio & Speech
Use cases: Meeting Notes · Translation & Subtitles · Audio & Speech · Automation
Deployment: Refer to project documentation
License: MIT

Best for

Developers, AI engineers, and researchers who need a multi-model speech processing toolkit to select and deploy specific ASR, VAD, punctuation, diarization, or emotion recognition models.

Key capabilities

Transcribes audio to text supporting various models for Chinese, English, Japanese, and up to 52 languages.
Processes audio in chunks for real-time transcription via WebSocket using Paraformer.
Identifies and segments speakers in audio using the CAM++ model.
Detects speech segments in audio using FSMN-VAD.
Adds punctuation to transcribed text using the ct-punc model.
Recognizes emotions in audio using SenseVoice or emotion2vec+large models.
Deploys an OpenAI-compatible /v1/audio/transcriptions endpoint via funasr-server.
Runs ASR as a single self-contained binary on CPU and edge devices without Python using llama.cpp and GGUF models.

Limitations and risks

Language coverage is checkpoint-specific and varies by model rather than being uniformly available across all configurations.
Pretrained model weights are licensed separately from the MIT toolkit source code. Adopters must verify licensing for each model used.
Operation requires downloading specific ASR models, such as Fun-ASR-Nano, SenseVoice, or Paraformer, before executing tasks.

Getting started

Installation difficulty is rated as easy because the toolkit installs via a simple pip command and provides prebuilt standalone binaries for CPU edge deployment. The quickest success path is running 'pip install funasr' followed by executing 'funasr audio.wav' in the CLI. No paid services are required.

Alternatives and comparisons

Provides an API server and Gradio WebUI for deploying multiple TTS, ASR, and voice cloning models.
A single model handling offline and streaming speech recognition, translation, and synthesis.

Project comparisons

Evidence and sources

README: Industrial speech recognition toolkit for offline, streaming, and edge deployment.
README: Whisper is a single model; **FunASR is a toolkit** — you pick the right model per job
README: Top speed | **340x realtime** (Fun-ASR-Nano + vLLM) | 13x realtime | ~1x realtime
README: - 2026/05/24: **vLLM Inference Engine** — 2-3x faster LLM decoding for Fun-ASR-Nano. Streaming WebSocket service with VAD + Speaker Diarization.
README: title={FunASR: A Fundamental End-to-End Speech Recognition Toolkit}

Reviewed at: 2026-07-17T12:02:14Z

Open original GitHub project