The-Speaker-s-Gym

Vocal variety arena (web preview)

Open the online app and choose Expand Your Vocal Variety from the menu, or use /#vocal-variety. It trains vocal variety plus structure: every round is a spoken answer to a real question using PREP (Point, Reason, Example, Final Point), while the game shows the speaker's voice live.

Seven games, all open from the start, grouped by what they train:

Game	What it trains
Volume: Loud & Soft	Guided ribbon. Boom the Point, whisper into the Example, land the Final Point strong.
Volume: Freestyle	Reach Whisper and Boom and swing between zones, your own way.
Pitch: High & Low	Guided ribbon. Lift the Point, paint the Example low then high, let the Final Point fall.
Pitch: Freestyle	Reach Low and High and swing between them.
Pace: Fast, Slow & Pause	Guided ribbon in words a minute: Slow for the Point, Fast for the Example, then pause and slow down.
Pace: Freestyle	Reach Slow and Fast and swing between them.
All Three: Full Vocal Variety	Swings in volume, pitch and pace in one answer, on three live lanes.

Pace is counted in words a minute, the same for every speaker: Slow is under 110, Steady is 110 to 139, Fast is 140 or more (pauses count as slowing down; the ribbon aims at 100, 125 and 150). Common published figures put conversation at roughly 120 to 150 and presentations at roughly 100 to 160, so 140 is an ordinary quick pace, not a strain. How the number is made steadier over a call: clips overlap by 0.8 seconds and each stretch of speech is counted from one clip only, so a word at a join is neither cut nor counted twice. The reading looks at about 8 seconds, and a long thinking pause counts for little more than an ordinary gap (a really slow speaker still reads slow). Words the speech service invents from silence or noise (it reports its own confidence, and "thank you" over almost no sound is a known invention) are removed on the server, and stretched or looping words are dropped in the browser. A failed clip is retried once, an expired sign-in is refreshed, and only a real refusal (404, 405, 501, a login that stays invalid, or four clips in a row failing) ends pace counting. A line under the pace reading says what is happening: how many words have been heard, or that the service is slow, or that no words have been heard (is the student unmuted, and is the Meet tab the shared one?).

Words are counted from the speech itself: a few seconds of audio at a time go to /api/transcribe (Whisper, the same service the AI feedback uses), which returns the words with their times. Readings appear a few seconds late on screen but are filed at the moment the words were spoken, so scoring is exact. If the service cannot be reached, pace falls back to an estimate from the sound alone (peaks in the vowel energy, wpmFromSound, fitted to recordings of known speed) and the results say so. Measured against recordings from 60 to 320 WPM, the sound-only estimate separates Slow, Steady and Fast only roughly, which is why words are counted whenever the service is available.

A round:

1. Read the question (swap it or have it read aloud) and pick the speech length: 1 minute, 1½ minutes or 2 minutes (default 2 minutes, remembered). The PREP parts keep their share of the time and the ribbon stretches with them; ramps at a PREP boundary keep their length. Longer speeches expect more swings. Each PREP part shows its vocal cue, such as "Point · Boom" or "Final Point · Low".
2. Prepare for 30 seconds to 2 minutes (default 1 minute, remembered). Optional notes for each PREP part. A sound check runs at the same time: whoever speaks now ("say a few words in your normal voice") sets the zones. "I'm ready" starts early and "+30s" adds time.
3. Speak for the chosen length after a 3-2-1 countdown. The current PREP part, its vocal cue and the student's own note are shown above the live stage, and a PREP notes panel under the stage shows all four notes with the part being spoken lit up ("No note" for a part left blank).
4. Results: stars, XP, a voiceprint ring (three rings for All Three), a map with the PREP parts marked, a part-by-part breakdown (which zone each part was mostly in, against its cue), range in the dimension's own units, playback, and a coaching tip.

Nobody is tuned. There is no "whisper, then shout" step. The zones follow whoever is speaking: the sound check sets the speaker's typical loudness (Whisper is more than about 6 dB under it, Boom more than about 5 dB over it) and natural note (six semitones either way reaches High and Low), the loudness carries on from one student to the next, and pace is absolute words a minute. This also covers microphones that are louder or quieter than usual and the levelling a video call applies.

Think Fast (reaction arena)

Pick Analogy Blitz ("A battery is like a jacuzzi, because…"), Hot Take (agree or disagree, and why), Sell It (pitch a strange product) or Wild Mix (all three). Choose 2 to 5 seconds a prompt and 5 to 20 prompts. All modes are open from the start. Hot Take (Toastmasters Table Topics style: pick a side, then Point, Reason, Example, Point) and Sell It (the classic "sell me this pen" test: who needs it, the problem it solves, a feature and its benefit, then ask for the sale) give 30 seconds a prompt, with a few seconds allowed for reading before the speed points start to fall; only Analogy Blitz uses the 2 to 5 second choice. Back goes one step: from a round or the results to the Think Fast menu, and only from the menu to Home. A 3-2-1 countdown, then each prompt shows a draining ring; start talking at once.

Hear me (a switch in Settings under Microphone, on by default; the games have no listening controls of their own) listens through the microphone, or through the meeting audio when that is on, and times how long it takes to start speaking. It learns the room's quiet during the countdown, so nobody is tuned or asked to shout. Each prompt scores up to 100: up to 80 for a fast start (full marks within 0.6 s, falling to 25 at the buzzer) and up to 20 for talking through the rest of the time. Answer 35 points or more to keep a combo going; the session score is the average, with stars at 35, 65 and 85. XP builds ranks (Rookie, Sprinter, Quick Draw, Lightning, Thunder, Mind Reader) and eight badges. Without a microphone it is a practice run: the prompts and timer only, no scores.

Sounds are skipped in meeting mode (they would be heard as answers). Nothing is recorded or uploaded. Progress is kept in this browser (sg-think-fast-v1), and sessions and prompts still add to the old Think Fast totals and the account sync. Files: think-fast-core.js (prompts, listening maths, scoring, ranks, badges; covered by tests/think-fast-core.test.js), think-fast.js and think-fast.css (screen), think-fast-words.js (analogy words).

Speak on the Spot

Get a question, think for a moment, then tell one story that answers it. New words appear every few seconds (whether or not you have used the last one) and stay on the board until you have worked them in; connect each into the story so it keeps flowing. Pick Quick Spark (30 s, 4 words), Classic (60 s, 6 words), Pressure (a new word every 6 s), Endurance (90 s, 10 words) or Custom. The question can be shuffled or typed in ("Write my own").

Speed is not scored. The score is out of 100 for the whole story: 45 points for how many of the words ended up in it (at any moment, even right at the end), 25 for how long it is (a full telling is about 90 words a minute of speech) and 30 for how well it connects. The connection is rated once, at the end, by the AI (api/story-score.js, signed-in users only); if it cannot be reached the connection is a capped estimate from joining words (because, so, then, but) and the results say "estimated". Saying a word before it appears does not count. Words are detected as before (the browser's speech recognition, or the call audio when Google Meet audio is on). Ranks, XP and eight badges are kept in this browser (sg-spot-v1). Rules live in spot-core.js (tested by tests/spot-core.test.js), the look in spot.css.

Playing over a video call (every game)

The coach shares their screen in a Google Meet and students speak from their own laptops. The games listen to the call instead of the microphone:

1. In Settings, turn on the one switch Use Google Meet audio in every game. It is off every time the app opens and is never remembered, so it only ever listens to the call when you have just switched it on. It is the only meeting control in the whole app: no game has its own, and there are no connect, reconnect or disconnect buttons.
2. Turning it on opens the browser's box. Choose your Google Meet tab and tick Also share tab audio, or (Windows) choose Entire Screen and tick Also share system audio, which also works when the call is in another window or app. If you cancel, the switch stays off. To pick a different tab, or to go back to your microphone, turn it off.
3. A live meter says ON: hearing the call. when a student speaks. If it stays flat, the panel says what to try (picked the wrong tab, shared without sound, a window, and so on). If the share ends (Stop sharing, or the Meet tab closes) the switch turns itself off and says so, and the games are on your microphone again.
4. While it is on, every game uses that one connection:
   • Speak on the Spot sends a few seconds of the call's audio at a time to /api/transcribe (with the target words as a spelling hint) and runs its existing word detection on the text.
   • Home-page recorder records the call without asking for the tab each time.
   • Vocal variety reads loudness and pitch from the call audio and counts words a minute for the pace games.
   • Think Fast times how fast each student starts, listening to the call (its Hear me switch in Settings decides whether it listens at all).
   • Warm Up does not listen to anyone, so it needs nothing.

Chrome or Edge on a computer only. The coach stays on the game tab (a round stops if the tab is hidden). Only the students are heard: the coach's own voice is not in the shared audio. Pitch and pace track well over a call; volume is the roughest, because the call levels and compresses voices before the game hears them, and each student's microphone settings apply as well.

Files: meeting-audio.js (connection, panel, ChunkTranscriber), api/transcribe.js (short speech to text), vocal-variety-game.js (rules, pure functions, covered by tests/vocal-variety-game.test.js), vocal-variety-core.js (pitch, vowel filter, sound-only pace estimate), vocal-variety.js and vocal-variety.css (interface).

Scoring. Each dimension maps to a 0 to 1 level. Variety = range (45%) + swings between zones (35%) + use of all three zones (20%). Guided games are 70% ribbon match and 30% variety; All Three averages the three dimensions. Talking for under half a round lowers the score; pauses are never penalised. Volume is relative level, not calibrated sound pressure. Strong means energetic, never strained.

Privacy. Audio and PREP notes stay in the browser. In meeting mode, and in the Pace games, a few seconds of speech at a time are sent to the app's speech service to be turned into words (nothing is stored by the app); the other games and the microphone-only Volume and Pitch games send nothing. Stars, XP, badges and the learned loudness are saved in this browser's localStorage only and never change the coaching scores, progress, or history. Capture stops on Finish, navigation away, backgrounding, Escape, or page close; the meeting connection stays open until it is disconnected (the browser shows its sharing indicator while it is).

api/transcribe.js only answers signed-in users: the browser sends its Firebase login token and the server checks it against the app's Firebase project (no secret needed, only the public project id). It is also limited to short audio chunks (1.5 MB) of known audio types. Every chunk with speech costs a little on the OpenAI account.

Web only. The loader and modules guard against native Capacitor, and build:web does not package them. Do not include them in an Android release until the user has tested and approved them.

Run npm test for the signal, rules and endpoint tests plus the existing app suite. The web deployment serves the repository root; www/ remains the Capacitor asset output. With a local server on port 8766 and Playwright CLI open on its index page, run playwright-cli run-code --filename scripts/validate-vocal-variety-ui.js to play the game with generated audio.