xfyun-chinese-asr-mcp

Node.js stdio MCP Server for Xingchen MCP cloud hosting. It calls Xunfei Chinese/English ASR large model and returns only recognition text.

MCP Tool

Tool name:

recognize_chinese_speech

Input:

{
  "audioBase64": "AAAA...",
  "sampleRate": 16000
}

Output:

{
  "text": "识别结果"
}

The audio must be raw PCM base64:

• 16 kHz or 8 kHz
• 16 bit
• mono
• raw PCM, no WAV header required
• short audio is recommended; keep each workflow call under 60 seconds

audioBase64 may include a data:audio/...;base64, prefix. The server strips that prefix before calling Xunfei.

Local Development

Install dependencies:

npm install

Run tests:

npm test

Build:

npm run build

Run locally over stdio:

XFYUN_APP_ID=your_app_id \
XFYUN_API_KEY=your_api_key \
XFYUN_API_SECRET=your_api_secret \
npm run dev

Xingchen MCP Cloud Hosting

Publish this package to npm, then create a cloud-hosted MCP Server at:

https://mcp.xfyun.cn

Use this startup configuration:

command: npx
args: -y xfyun-chinese-asr-mcp

Environment variables:

XFYUN_APP_ID=your_app_id
XFYUN_API_KEY=your_api_key
XFYUN_API_SECRET=your_api_secret

After hosting succeeds, add the MCP tool in the Xingchen workflow Agent intelligent decision node. The workflow should pass the board-side raw PCM base64 into audioBase64 and read only the returned text.

Board-Side Contract

The board should send raw PCM base64, not local file paths:

{
  "audioBase64": "AAAA...",
  "sampleRate": 16000
}

If the board captures WAV, remove the WAV header or encode the PCM payload only. If the board captures another format, convert it to 16-bit mono PCM before calling the workflow.