Speech practice you can check

Say it. See it. Say it better.

Record a short talk. Get numbers you can verify — pace, filler words, pauses, pitch variation — with the evidence behind each one, then choose one thing to change. No grades, no judging your voice or accent, nothing stored on our server.

Try the Studio Students: build your own

One real plane. One model plane.

This project runs on two lines at once. The full-size plane flies today. The model plane is built by a student, part by part, until it flies too.

Line 1 · the full-size plane

Speech Lab Studio live

Built by LIGHT HOPE to production standards: on-device measurement, opt-in speech recognition and structure feedback, every AI claim checked against the transcript, privacy by default. Ready for classrooms, clubs and coaches.

  • Works fully without AI — the measurements run in your browser
  • AI steps only after you tick a box, for each kind of data
  • Model quotes are verified word-for-word; unverified ones are flagged
Open the StudioFor schools

Line 2 · the model plane

Build your own speech coach student kit

Eight stages, each one a real part of the plane. Write each function yourself, run the tests in your browser, keep a build log, then assemble your own app in your own repository. Every line is yours.

  • Tests tell you what is still broken — never the answer
  • Hints come in steps; you decide how many to open
  • Your code, your log, your decisions — exportable any time
Start building家长说明

The parts of the plane

Each stage of the student kit builds one part. The same ideas — not the same code — power the Studio.

Airplane parts diagram A plane seen from above. Numbered labels mark eight parts: 1 fuselage, 2 airspeed indicator, 3 drag check on the wings, 4 rudder at the tail, 5 flight log, 6 black box, 7 engine, 8 test flight. 1Fuselage 2Airspeed 3Drag check 4Rudder 5Flight log 6Black box 7Engine 8Test flight
  1. Fuselage — count the words. Every other number rests on this one.
  2. Airspeed — words per minute. A rate: change in words ÷ change in time.
  3. Drag check — filler words. Why “umbrella” must not count as “um”.
  4. Rudder — pauses from loudness. Frames × frame length = seconds; boundaries matter.
  5. Flight log — did I get better? Current − previous, and when a comparison is fair.
  6. Black box — check the evidence. A quote that exists still isn't proof the judgment is right.
  7. Engine — your AI step. Designed on paper first: what goes in, what checkable thing comes out.
  8. Test flight — your app. Assemble the parts in your own project, record five times, keep the evidence table.

The practice loop

1 · Record

30–150 seconds. Or upload, or try a sample.

2 · Check

Read the transcript. Put back every “um” the machine dropped.

3 · See

Numbers with the method behind each one. Missing data shows “—”, never 0.

4 · Change one thing

Pick one. Record again. Compare only the same task.

What we promise

  • No accounts. No audio, transcript or result is stored on our server.
  • AI is opt-in, per kind of data, every time.
  • We never score the person: no confidence, emotion, accent or “speaking ability” ratings.
  • Every number shows how it was made.

What we have not proven yet

  • That the feedback improves anyone's speaking — that needs a real pilot.
  • How well speech recognition keeps “um” across many voices and microphones.
  • Behaviour on every browser (tested: desktop Chromium; Safari/iOS still to check).

We write down what we have not tested, instead of guessing. How we test.