FunASR Agent Integration — Give Your AI Agent Ears

Quick Start

Two lines to add speech recognition to your agent stack:

# Install & start
$ pip install funasr fastapi uvicorn python-multipart
$ funasr-server --device cuda --port 8000

# Now any OpenAI SDK client works
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
result = client.audio.transcriptions.create(
    model="sensevoice", file=open("audio.wav", "rb")
)
print(result.text)
            

Integration Methods

REST API

OpenAI-Compatible Server

Drop-in replacement for /v1/audio/transcriptions. Works with LangChain, AutoGen, CrewAI, Dify, Flowise, Open WebUI — any framework using OpenAI audio API.

funasr-server --device cuda --port 8000

MCP

Model Context Protocol

Add FunASR as a tool in Claude Code, Cursor, or Windsurf. Your AI assistant can transcribe any audio file directly.

// claude_desktop_config.json
{"mcpServers": {"funasr": {
  "command": "python",
  "args": ["funasr_mcp.py"]
}}}
                    

WebSocket

Streaming Server

Real-time voice input via WebSocket. 2Pass mode: instant partial results + high-accuracy final correction. For voice agents that need low latency.

docker run -p 10095:10095 \
  funasr-runtime-sdk-online-cpu
                    

Python

Direct SDK

Use FunASR directly as a Python function in your agent code. No server needed for single-process applications.

from funasr import AutoModel
model = AutoModel(model="iic/SenseVoiceSmall")
text = model.generate(input="a.wav")[0]["text"]
                    

Compatible Frameworks

Framework	Stars	Integration	Method
LangChain	137K	OpenAI audio tool	Change base_url
Dify	142K	STT provider	OpenAI-compatible endpoint
Open WebUI	138K	Speech-to-text	OpenAI-compatible endpoint
AutoGen	47K	Agent tool function	OpenAI SDK
Flowise	30K	STT node	OpenAI-compatible endpoint
Claude / Cursor	—	Audio transcription tool	MCP Server
Pipecat	12K	STT service	WebSocket / OpenAI
LiveKit Agents	—	STT plugin	WebSocket streaming

Available Models

Model	Speed (GPU)	Speed (CPU)	Languages	Best For
sensevoice	170x realtime	17x realtime	zh/en/ja/ko/yue	General + emotion
paraformer	120x realtime	15x realtime	zh/en	Chinese production
fun-asr-nano	17x realtime	3.6x realtime	31 languages	Multilingual + LLM

Get Started

Two commands. No config files. No Docker required.

pip install funasr fastapi uvicorn python-multipart
funasr-server --device cuda
            

OpenAI API Docs MCP Server Docs