FunASR Speech Recognition Plugin

Self-hosted speech recognition through FunASR's OpenAI-compatible API. The plugin supports multilingual ASR models for private Dify deployments without sending audio to a hosted transcription service.

In FunASR's 192-minute benchmark, SenseVoiceSmall reached 170x real-time on GPU and 17x real-time on CPU. Whisper-large-v3 reached 13x real-time on GPU in the same benchmark. Results depend on the hardware, audio, model, and concurrency settings.

Setup

1. Deploy a SenseVoice server:

pip install "funasr>=1.3.26" fastapi uvicorn python-multipart
funasr-server --device cuda --model sensevoice --host 0.0.0.0 --port 8000

Use --device cpu when CUDA is unavailable. Fun-ASR-Nano deployments use vLLM and can be installed separately:

pip install "funasr>=1.3.26" vllm fastapi uvicorn python-multipart
funasr-server --device cuda --model fun-asr-nano --host 0.0.0.0 --port 8000

2. In Dify, configure this plugin with the server URL, for example http://your-server:8000 or http://your-server:8000/v1. The plugin normalizes both forms to a base URL ending in /v1, such as http://localhost:8000/v1. Leave the API key empty when the server does not require one.

Predefined models accept audio files up to 25 MB.

Supported Models

• sensevoice - Mandarin, Cantonese, English, Japanese, and Korean ASR with emotion and audio-event tags (default)
• paraformer - Chinese ASR with punctuation
• paraformer-en - English ASR
• fun-asr-nano - Chinese, English, Japanese, and Chinese dialect/accent ASR with an encoder + LLM architecture