Specifications · Evaluation Engine

Technical specs · for engineering leads

Two contiguous sections: dual-mode evaluation (streaming + file) and the full assessment parameter map (four main dimensions + phoneme-level alignment + diagnostic types). Raw API fields, JSON shapes and stream protocols.

Dual Mode

Record-and-Evaluate · Dual Mode

Streaming real-time evaluation + file-based evaluation, covering interactive practice to batch QA.

Real-time Recording Evaluation

Capture voice from microphone and stream via WebSocket to the engine. No temp files, instant scoring, ideal for interactive speaking practice.

Evaluation Flow

Create a stream session with eval type and reference text

Start recording and stream audio to cloud evaluator

Stop recording and get multi-dimensional scores instantly

30-50% lower latency, smoother UX
No need to manage audio files
Auto reconnect on network drops

Audio File Evaluation

Supports local path, Base64, and URL. Pass a path and the proxy handles encoding/upload automatically, ideal for batch workflows.

Three Input Methods

audio_file_pathLocal path, easiest

audio_base64Base64 data

audio_urlRemote URL

Supports mp3/wav/ogg/m4a/aac
Auto-compress large files
Great for batch evaluation and playback analysis

Latency

30-50%↓

vs. traditional upload-back flow

Audio Formats

5 types

mp3 / wav / m4a ...

Input Methods

Path / Base64 / URL

Auto Reconnect

Resume on drop

Friendly in weak networks

Task Coverage

WordwordSentencesentParagraphparaSemi-open5-dim scoringOpen questionPicture speaking / essayFree recognitionRealtime ASR + punctuationAI TalkHuman-AI dialogue

Assessment Parameters

All Assessment Parameters

The MCP interface mirrors Chivox's exam-grade engine API: four main scores, phoneme-level alignment, and error classification — all LLM-ready.

Overall

Weighted composite of pronunciation, integrity, fluency and rhythm

Accuracy

Pronunciation accuracy down to the phoneme level

Integrity

Detects insertion, omission and skipping — did they actually finish?

Fluency

Naturalness of flow and pause placement

Rhythm

Stress, prosodic phrasing and intonation combined

Speed

Live WPM / SPM — penalised for both too fast and too slow

Diagnostics

Per-word / per-phoneme tagging: normal / omit / insert / mispron

Phoneme Alignment

Score, timestamp and confidence for every phoneme

mcp-response.json

{
  // 四大主维度
  "overall": 72,
  "pron": {
    "accuracy":  65,
    "integrity": 95,
    "fluency":   85,
    "rhythm":    70
  },
  "speed": 130 // WPM,

  // 词/音素级诊断
  "details": [
    {
      "char":  "record",
      "score": 58,
      "dp_type": "mispron",
      "phonemes": [
        { "char": "r", "score": 45, "dp_type": "mispron" },
        { "char": "ɪ", "score": 78, "dp_type": "normal"  }
      ]
    }
  ]
}

dp_type: normal / omit / insert / mispron考试级 API

Hover any field on the left to highlight the mapped JSON key on the right