01 · Choose the boundary
Decide where scoring enters the product.
Start from who captures audio, who decides when to score, and where the structured result needs to land. That usually makes the integration choice obvious.
/integrations
Three doors into the same engine.
Whether your product records the audio, your IDE agent calls a tool, or your own orchestrator decides when to score — the request carries the same reference text and the response carries the same phoneme-level detail.
/side-by-side
Same request. Same response.
Switch between the three surfaces below — the reference text goes in the same way and the scores come back in the same shape.
Your app records
SDK & API
Send the recording with reference text. The engine aligns it against native speech models and returns evidence your product can use.
engine.start({
coreType: 'en.sent.score',
refText: 'May joy and health be with you always.'
});iOS · Android · Web · serverWebSocket streamingWord and phoneme detail
Explore SDK & APISame structured response
Scores, not just a transcript.
AudioChivoxJSON
Accuracy82
Fluency85
Integrity95
Tone88
Includes word, phoneme and Mandarin tone evidence when the assessment mode supports it.
/pick-one
Which one fits your build?
| SDK & API | MCP server | Function calling | |
|---|---|---|---|
| Best for | Apps that capture the audio themselves | LLM clients and agent IDEs | Agent stacks you already orchestrate |
| Runs in | iOS · Android · web · server | Cursor · Claude Desktop · any MCP client | Any backend or agent framework |
| Transport | WebSocket streaming and REST | Streamable HTTP | REST or WebSocket |
| Auth | Application ID + signed timestamp | Bearer API key | Bearer API key |
| Who decides the call | Your code | The model, from the tool list | The model, from your schema |
| Commercials | Licensed via BD | Evaluation points, self-serve | Evaluation points, self-serve |
| Best first milestone | Score one real recording inside your product UI | Let an agent select and call the right assessment tool | Complete one tool call inside your existing agent loop |
| Your team owns | Capture UX, task design, thresholds and learner feedback | Agent instructions, permissions and how results are presented | Tool routing, validation, retries and response handling |
Still unsure? Teams building a learner-facing app commonly start with the SDK, while teams adding assessment to an agent commonly start with MCP. The best choice is the one that keeps audio capture, product logic and operational ownership in the right part of your stack.
/pilot-to-production
Prove the workflow before you scale the traffic.
A strong pilot answers a product question, not only an API question. Use one real task to validate scoring, feedback, failure handling and operating expectations together.
02 · Prove one task
Use representative audio, not a perfect demo clip.
Test the language, device mix, accents and background conditions your users actually bring. Review both successful scores and rejected-audio paths.
03 · Set product rules
Turn evidence into one clear next action.
Agree thresholds, retry rules and learner-facing feedback before scaling the integration to more task types or agent behaviors.
04 · Prepare production
Confirm traffic, privacy and support expectations.
Separate environments, scope keys, add spend controls, and align on retention, procurement and service requirements before launch.
Planning security, limits or monitoring at the same time?Review runtime and operations
/contact
Let’s build a better speaking experience together.
Tell us what you’re building and whether you need SDK/API, MCP, or function calling. We’ll reply within one business day with pilot credits, pricing, or a deployment plan.