Speech On

A free guided speech warm-up at speechon.app. Open the link and practise. No account, no sign-up, no microphone, no backend, no build step, no runtime dependencies.

Speech On is not a medical device and is not a treatment for any condition. It is a general practice and wellbeing tool. See /terms, /safety and /for-clinicians.

What it does

Pick 5, 8, 12 minutes or the full sequence (~20 minutes, 18 exercises), pick an occasion, and work through guided exercises: settling the breath, relaxing the face, jaw and lip and tongue movement with speech attached, deliberate articulation, word endings, consonant clusters, rate control and pausing, prosodic emphasis, connected reading, and transfer into unscripted speech.

Every exercise carries a written caution. Every step is self-paced; an optional timer can be switched on in settings and measures nothing. The shorter routines are subsets of the same exercises.

The complete routine is readable without starting a session at /routine, which is generated from content.js so it cannot drift from the app.

The daily three: Speak, Read, Think

The home page shows today's three practices and their streaks: a full-day streak (all three done), per-practice streaks, a month calendar, and two calendar files generated on the device (a repeating daily reminder for all three, and practised days as all-day events).

• Read (/read), about 10 minutes. One paragraph at a time: read, the text hides, explain it from memory, check against a model, note whether attention drifted and what got in the way, link it to the previous paragraph. Then three points from memory, three questions and two vocabulary words that return on a spaced schedule (1, 3, 7, 16, 35, 90 days). The passages are 47 verbatim excerpts from public-domain classics (Smith, Mill, Madison, Hobbes, Machiavelli, Thucydides, Keynes, Plato, Aristotle, Bacon, Emerson, Thoreau, Du Bois, Tolstoy, Wilde, Ruskin, Babbage, Taylor and others) across business, philosophy, politics, history, art and technology, taking turns by category, with a context note and glosses for older words. "Read a real article" runs the drill on a piece you read on the publisher's own site (FT, The Economist, Aeon, Foreign Affairs and others): nothing is copied into the app. "Use it on your own article" runs the same moves on pasted text, which never leaves the browser.
• Think (/think), about 10 minutes. An energy and sleep check-in, a 12-problem business-maths sprint generated across eight adaptive levels (percentages to EV/EBITDA and dilution, with earlier levels interleaved), one Fermi estimate with a worked chain, and "Say it in 60": a structured spoken answer, out loud, with a self-check. Not brain training: it practises the specific skills, and says why on the page.

Neither page assesses, diagnoses or treats ADHD, dyslexia or anything else, and both say so.

Design commitments

• No microphone. There is no code that could record audio, and the site sends Permissions-Policy: microphone=() on every response.
• No third-party requests. Fonts are self-hosted. No CDN, no tag manager, no pixels, no cookies.
• Nothing leaves the browser. Practice history, preferences and the unfinished session live in one localStorage key. There is no server to send them to.
• Analytics are optional and anonymous. Vercel Web Analytics is cookieless, honours Do Not Track and Global Privacy Control automatically, and can be switched off on /privacy.
• Works offline once loaded, including the reference pages.
• No claims it cannot support. The routine has not been clinically validated. The face massage and lip trill are labelled on screen as a settling step and a voice warm-up respectively, not articulation exercises.

Files

File	What it is
index.html	App shell: home, player and completion screens
app.js	Session engine, state, streaks, pacer, data export and import
content.js	Exercise structure, routes, artwork labels and the English pack
lang/*.json	Translated packs, loaded on demand
art.js	Articulation artwork: front-on mouth shapes per exercise
styles.css	App styles, light and dark, self-hosted fonts
page.css	Shared styles for the static reference pages
routine.html	Every exercise in full, generated from content.js
safety.html terms.html privacy.html accessibility.html	The reference pages
for-clinicians.html handout.html	Clinician page and a printable patient handout
trio.js train.css	The daily three: today's ticks, streaks, month calendar, calendar files, iPhone install hint
read.html read.js	Read: guided reading practice
reading/classics.json	47 verbatim public-domain passages with their questions; reading/index.json holds ids, titles and the reading order. tools/classics/ rebuilds them
think.html think.js	Think: maths sprint, estimate, Say it in 60
think/*.json	344 estimates and 368 speaking prompts, built from tools/*_src.txt
tools/	build_think.py builds the Think banks; build_reading_index.py validates every passage and writes the index
sw.js	Versioned offline shell, network first
vercel.json	Clean URLs, redirects, cache and security headers

Languages

The interface and every exercise are translated per language; the exercises are adapted rather than translated word for word, because sound contrasts do not carry across languages. Safety, legal and About text stays in English in every language, deliberately: a mistranslated caution is worse than an untranslated one, and the app says so.

Only languages with a pack in lang/ appear in the picker (ready: true in content.js). Anything a pack leaves out falls back to English rather than going blank.

Shipping now: English, Spanish, French, German, Portuguese, Indonesian, Hindi, Kannada. Queued: Russian, Arabic, Chinese, Japanese, Bengali, Marathi, Telugu, Tamil, Gujarati, Malayalam, Punjabi, Odia, Urdu.

Every translation except English is machine-assisted and has not been reviewed by a native speaker. The app says this to the user on the home screen. Native-speaker review of any pack is the single most useful contribution anyone could make; open an issue.

Adding a language

1. Copy lang/es.json and translate it. Keys the app needs are in content.js under EN.ui, EN.contexts and EN.steps.
2. Adapt the exercise material rather than translating it: the sound pairs, word endings and consonant clusters must be contrasts that exist and matter in that language.
3. Add the language to LANGUAGES in content.js with ready: true.
4. For a right-to-left language set dir: 'rtl'; the layout and the arrow-key navigation already handle it.

Accessibility

Target is WCAG 2.2 AA; the app is partially conformant and the known gaps are listed at /accessibility. Keyboard operable throughout, screen-reader labelled, reduced motion honoured (the artwork shows both positions side by side instead of alternating), dark mode, a text-size control, and contrast of at least 12:1 for body text in both themes. Checked down to 320 CSS pixels.

The largest known gap is that there are no audio recordings of the sounds and phrases, so a user who cannot read the material has only synthesised speech. That is the next thing to build.

Run locally

python3 -m http.server 8765

Then open http://localhost:8765. Serve over HTTPS in production for the installable app, clipboard, sharing and the service worker; all of these degrade gracefully without it. Analytics never load on localhost or over plain HTTP.

Editing content

Passages live in reading/r*.json (one object per passage: title, topic, purpose question, paragraphs, model gists, links between paragraphs, three key points, three questions with the correct option first, two vocabulary items whose word must appear in the text). Estimates and prompts are one line each in tools/estimates_src.txt and tools/prompts_src.txt. After any edit:

python3 tools/build_think.py && python3 tools/build_reading_index.py

The validator fails loudly on malformed passages, duplicate titles or a vocabulary word missing from its passage.

Deploy

Vercel static site from the repository root: no framework, no build command, no output directory. vercel.json handles clean URLs, the fluency.html redirect, cache headers and the security headers. Increment CACHE in sw.js whenever the shell changes, then test both a returning visitor and a fresh browser.

Contributing

Most useful, in order: native-speaker review of a language pack; a recorded reference voice for any language; review of the routine by a registered speech and language therapist; accessibility testing with real assistive technology. Corrections to anything on /routine or /for-clinicians are very welcome — open an issue.

A project by Vyshak K Iyengar.