ChivoxMCP
Sign in
liveThe listening layer for voice-native agents

Speech & pronunciation assessment MCPYour agent can hear them.Now it can grade them.

Chivox MCP turns raw speech into a dense, agent-ready payload — phoneme scores, stress, tone and fluency in one MCP call, ready for any LLM.

Start freeSee it run
Deep linguistic understanding
Enterprise-ready
Real-time intelligence
54321你好T3 ✓T3 ✓nǐhǎoT3 + T3 → T2 + T3TONE SANDHI · DETECTEDSCORE92thinkPHONEMEACCURACYHEARD/s/WEAKDROPPED—/s//θ//ɪ/✓/ŋ//ŋ//k//k/3 ISSUES · DETECTEDMISSING · WEAK · MISPRONOUNCEDALL CORRECTEDPHONEME DIAGNOSIS · SUPERVISEDSCORE5892
/product highlights5 frames
01
One MCP. Every agent runtime.
Plug Chivox into Claude, Cursor, Cline, LangChain, or any custom loop in minutes.
One npx command — no SDK to install.
Same payload for Mandarin and English.
Works with any MCP-compatible client.
01 / 05Plug-and-play
/the-feedback-loop

From speech to the next best practice

Chivox handles the acoustic judgment. Your LLM receives structured evidence it can explain, reason over and turn into the learner’s next action.

The feedback path
  1. 01/Speech in
    Capture the learner
  2. 02/Evidence out
    Return acoustic detail
  3. 03/Next action
    Let the agent respond
Speech score meters: overall 84, accuracy 78, fluency 88, rhythm 73
/assessment

Score guided speech

Stream live audio or post a file. Get overall, word and phoneme-level evidence in one response.

signals
accuracyfluencyphoneme
AI dialogue scoring UI with five-dimension score chips
/conversation

Evaluate open dialogue

Score free-flow responses across fluency, content, grammar, accuracy and rhythm — turn by turn.

mode
AI-talk5-dimstreaming
Bilingual panel with zh-CN 你好 / en-US Hello pronunciation details
/language depth

Diagnose English and Mandarin natively

Inspect tones and pinyin in Chinese; stress, rhythm and CEFR-aligned evidence in English.

coverage
zh-CNen-USCEFR
Personalized drill card with /θ/ minimal pairs and LLM chips
/agent outcome

Turn evidence into the next practice

Give the structured JSON to any LLM to coach, route or generate targeted drills for the next turn.

works with
GPTClaudeGemini
/evidence-you-can-inspect

Acoustic depth you can inspect. Scale you can trust.

Twenty years of speech-assessment R&D, exposed through one stable contract. Toggle zh / en to inspect the same pron.* / details[] structure; use the benchmarks beside it to sanity-check Chivox against your own evaluation harness.

Mandarin · tone accuracy

你好，今天天气……

nǐ hǎo, jīn tiān tiān qì
78/100
sentence score
你
nǐ
T3
85
好
hǎo
T3
72
今
jīn
T1
88
天
tiān
T1
88
天
tiān
T1
58
气
qì
T4
91
tonesT1T2T3T4
LLM hint · second 天 (tiān)collapsed into T4. Keep the pitch high and steady — it’s a T1.
95%+
agreement with human experts
r ≈ 0.95

Scores align with certified human expert rubrics at 95%+ correlation. Validated by national standardized speaking tests in 100+ cities.

0.95+
Pearson r vs experts
<2 pts
Mean absolute error
500K+
Calibration utterances
  • Per-dimension rubrics: pron, fluency, completeness, prosody.
  • Calibration corpus refreshed quarterly across L1/L2 cohorts.
  • Stable across mic quality, room noise and child voices.
Validated against national speaking-test rubrics · ISO/IEC 17025-aligned labs
/quickstart

Get the first structured score in 3 steps

Paste the config, connect Chivox, then call one assessment tool from your agent loop.

Full docs & API reference

Grab an API key

Sign up, confirm your email, copy the key. Free trial credits included.

Get a key
02

Add one block to your MCP configrunning

Paste the snippet into Cursor, Claude Desktop, or your custom agent — pick a tab on the right.

03

Call a tool from your LLM

Hand your model the audio. It gets back nested JSON: pron sub-scores, fluency + WPM, audio SNR, and details[] with ms ranges, stress, liaison and per-phoneme rows.

API reference
Live playground · no micRun a real Mandarin + English demoWatch raw JSON → teacher diagnosis → auto-generated drill. No signup, no setup.Open the playground
$
npx -y @chivox/mcp
/product paths

Start with the learner you are building for

English, Mandarin and Kids share one MCP contract, while each product path gives your agent the language and learner context it needs.

English learner with overall score 84, fluency 78, and /θ/ pronunciation tip
English assessment

Pronunciation, fluency and phoneme feedback

Score English speech with explainable dimensions — so tutors and agents can coach the exact sound, not a black-box percentage.

Explore product
Learner practicing Mandarin tones with pinyin chips and score 88
Mandarin assessment

Tone, Pinyin and fluency scoring

Give agents tones, sandhi and phoneme-level Mandarin — acoustic detail a transcript-only stack cannot surface.

Explore product
Young learner unlocking a star after pronunciation practice on a tablet
Kids speech assessment

Structured feedback for young learners

Keep the raw scores behind the scenes; surface one clear next step so practice stays encouraging and age-appropriate.

Explore product
PricingPay for results

Simple points for successful speech evaluations.

One point for a word or sentence. Two for a paragraph. Failed calls use zero points, so you only pay when Chivox returns an assessment.

  • Successful evaluations only
  • Shared across every API key
  • Points stay valid for 30 days
Start with 600 free points
How points work

From free trial to production

No card required
  1. 1

    Start free

    Valid for 30 days

    600 pts
  2. 2

    Evaluate successfully

    Points are deducted only when an assessment returns.

    word · sentence−1 pt
    paragraph−2 pts
  3. 3

    Top up as you grow

    Higher packs lower your unit cost.

    +20%
Failed calls cost $0 and use 0 points.
Ready to wire it up?

Same payload. Your agent. Your production loop.

Drop Chivox MCP into Cursor, Claude Desktop, or any agent SDK. One npx and you’re reading the same JSON you just saw above.

Free trial · spend caps · low-balance alerts · zero audio retention

See quickstartRead the docsGet your API key