ChivoxMCP
Back to home
Product Value

Three Core Values

From fine-grained scoring data to deep LLM analysis and production-ready deployment — one protocol unlocks AI listening and speaking capability.

01

Rich Data Dimensions

Four core dimensions (Accuracy · Integrity · Fluency · Rhythm) + phoneme alignment + error typing, all LLM-ready.

MCP Response (partial)live
{
  "overall": 72,
  "pron": {
    "accuracy":  65,
    "integrity": 95,
    "fluency":   85,
    "rhythm":    70
  }
}
Error Types · dp_type
normalmispronomitinsert
Accuracy
Integrity
Fluency
Rhythm
Diagnostics
Phoneme Alignment
02

LLM Second-Pass Diagnosis

One clear prompt turns Chivox MCP JSON—scores, phonemes, dp_type—into classroom-ready diagnosis and drills. Start from reusable prompt-skill templates.

System prompt · prompt skill
Group issues by dp_type (e.g. mispron, omit, insert). For phonemes below score 70, spell out articulation cues (lip, jaw, tongue) and rank what to practice first.
↓
LLM output

/ð/ realized as /z/: light tongue tip on the back of the upper teeth—not an alveolar fricative. /ɒ/ too tight: lower the jaw a little and relax the tongue root. Prioritize /ð/; it is extremely common in English.

Works with common frontier models
GPTClaudeGeminiQwenDeepSeekGLM
  • Cluster weak phonemes → actionable correction steps
  • Level estimate + concrete gaps to reach the next band
  • Personalized drills aligned with the learner profile
03

Power Any AI Product

Based on MCP, any compatible AI client gains exam-grade speech evaluation with one line of config — no SDK wrapping required.

~/.cursor/mcp.json1-line
{
  "mcpServers": {
    "chivox": {
      "command": "npx",
      "args": ["@chivox/mcp"]
    }
  }
}
MCP-native Clients
Cursor
Claude Desktop
扣子 Coze
Dify
豆包
LangChain
Scenario Integrations · IM Bots
Feishu
DingTalk
WeCom
  • One-line integration
  • Platform-agnostic
  • Boundless scenarios
All three layers can be invoked directly by LLMs via MCP tools