Skip to content

Translation plan

Regis edited this page Nov 14, 2026 · 6 revisions

Translation plan (analysis)

The owner's next project (21 October 2026): the web console, the statistics page, graphs1090's page, and the Bluesky posts and their pictures in French, German, Dutch, Indonesian, Spanish, Portuguese (Brazil), Russian, Chinese, Korean, Japanese and Ukrainian (and Italian, added with the owner's answers), chosen at the console's first sign-in or from flags at the top; the Linux menu stays in English. And, in the posts, your own message assembled with ready-made ADS-B tags. This page is the analysis asked for first: is it feasible, what can break, what it costs the Pi.

Where it stands: the web pages are translated — the console, the statistics page and graphs1090's page: French in 0.47.0, eighteen more in 0.48.0, Swahili in 0.53.1 (published by the owner before a native review). The language is asked at the console's first visit and chosen from a flag at the top of every page. The posts (phase 3 below) are next. How to correct a language: Translating.

The owner's answers (22 October 2026, 00:01)

  • Translate everything into all the languages covered, on the console, the statistics and graphs1090, from now on: every new or changed English string goes into every language in the same release.
  • The translation reference is corrected when a suggestion from abroad proves inexact: a reviewer's correction is checked against the English before it is taken (tools/i18n.py apply --dry), and one found inexact is not (--skip); a translation of ours found wrong is corrected in the catalogue and its sheet.
  • The console's first page not crowded with countries: the browser's language and English as two buttons, every other language in one list.
  • More languages: Finnish, Estonian, Latvian, Lithuanian and Polish (Latin script: with German and the others), and Swahili, Arabic, Hausa and Yoruba, except those whose translation could not reach 90%. Hausa and Yoruba are not included: Yoruba is written with tone marks on most vowels, which a machine translation gets wrong too often for a reader to trust, and neither has the settled technical vocabulary (settings, feeders, receiver, decoder) that the pages need; both can come later from a native speaker's sheet. Arabic is in. Swahili was translated but held back: its translator could not promise 90% (estimated 85 to 92%), the technical words differing between Tanzania, Kenya and the software glossaries. The owner published it on 22 October 2026 (0.53.1); a native speaker's review of its sheet is still welcome.
  • A picture for a language of many countries: Swahili shows the map of Africa; Arabic, spoken from Morocco to Iraq, shows its letter ʿayn (ع) rather than one country's flag. Finnish, Estonian, Latvian, Lithuanian and Polish their countries' flags. Arabic is right to left: the console's side bar, drawer and switches mirror; what the Pi prints, commands, numbers and the charts stay left to right; numbers in Western digits, as the Pi writes them.

The owner's answers (21 October 2026, 23:06)

  • Chinese: both, Simplified (zh-Hans, China's flag) and Traditional (zh-Hant, Taiwan's flag).
  • English: the two flags merged, the United Kingdom's and the United States', cut along the diagonal.
  • No unreviewed mark: instead, per language, a document with the English as the reference where people write their local corrections, which the owner puts back: the review sheets docs/translations/<lang>.csv and tools/i18n.py apply (Translating).
  • The terminal output stays in English: yes.
  • Italian too.

In short

  • Feasible, with three limits that are physical, not choices: graphs1090's graph images stay English (they are drawn by graphs1090's own script); the dot-matrix and split-flap pictures can carry accents and Cyrillic once their 5 × 7 font learns them, but no Chinese, Japanese or Korean (a 5 × 7 cell cannot draw them: their labels stay in English, the post's text and the summary cards are translated); the terminal output shown in the console's cards (what zenithboard prints) stays English, as the Linux side does.
  • Load on the Pi: negligible, except one optional package: Chinese, Japanese and Korean summary cards need the Noto CJK font (61 MB to download, 91 MB on the card, +4.4 MB of memory while a card is drawn, measured), installed only when one of those three is chosen for the posts.
  • Regressions: kept in check by the English text staying the key (gettext style), a test of every catalogue, and previews in three scripts (French, Russian, Japanese) at desk and phone width before each merge.
  • Translations are machine-made (by the session that builds them), and each language has a review sheet where a native speaker corrects them (the owner's choice over an unreviewed mark). Aviation words (callsign, squawk, MLAT) are checked against each language's usage, not guessed.

What is to be translated (measured, 0.45.0)

Part Where Strings Words (English)
Web console console.js, console.html, labels in console_actions.py, server answers in console.py about 600 about 4 500
Statistics page stats.js, stats.html about 80 about 300
graphs1090's page zenithboard.js (its buttons and panel titles, set by script) about 30 about 60
Posts, summaries, their pictures social.py, social_week.py, social_card.py about 300 about 3 000
Total, per language about 1 000 about 8 000

Eleven languages: about 88 000 words to write, and to review. Country names (about 250, in the lists, the flags' captions, the posts) come from the Unicode CLDR data the browsers carry (Intl.DisplayNames) on the pages, and from a table generated once from the same data for the posts (about 60 KB for twelve languages; CLDR's licence is permissive). Airline, model and city names are proper names and stay as they are.

How (the design proposed)

One catalogue per language and part, flightinfo/i18n/<lang>/console.json, stats.json, social.json; the English text is the key: t("Make a backup"). A missing translation shows the English, never a blank; the tests grepping English strings keep working; the English page is byte for byte what it is today.

Plurals and numbers: the browser's Intl.PluralRules, Intl.NumberFormat, Intl.DateTimeFormat on the pages (Russian and Ukrainian have three plural forms, Chinese, Japanese and Korean none); a small table of the same rules in Python for the posts.

Choosing the language

  • The console's first sign-in: a first screen before the nine first steps, the languages by their own names with their flag; written to the Pi as CONSOLE_LANG (the default for every browser) and to the browser.
  • At the top of every page (console, statistics, graphs1090): a flag with the language's own name (Français, Deutsch, 日本語), a list to change it. Flags are pictures (the flags ZenithBoard already carries), not emoji: Windows draws no flag emoji. A flag names a country, not a language: English shows the UK's, Portuguese Brazil's (as asked), Spanish Spain's, Chinese China's (simplified characters, unless the owner prefers traditional).
  • The posts: SOCIAL_LANG, asked in the terminal's Bluesky set-up (social setup) and in the console's Sharing.

Your own message, with tags (in the posts' rotation as one more style, custom, beside factual, story, board, spotter, lookup): one or more templates written in your words, in your language, with tags that ZenithBoard fills in. A part between [ and ] is left out when one of its tags is empty. Example:

{flag} {airline} {flight}[, {from} → {to}] · {type}[ {registration}] at {altitude}. {why} {hashtags}
Tag Means Example
{flight} The flight number when known (IATA), else the callsign FR4512
{callsign} The callsign as transmitted RYR4512
{airline} The airline's name Ryanair
{airline_country} The country the airline is based in Ireland / Irlande
{flag} That country's flag, as an emoji (Bluesky shows it) 🇮🇪
{registration} The aircraft's registration EI-DCL
{reg_country} The country it is registered in (by its address) Ireland
{type} Make and model Boeing 737-800
{type_code} ICAO type designator B738
{category} What it is: airliner, cargo, business jet, helicopter, military, light aircraft… airliner
{from} · {to} Departure and arrival, city and code Dublin (DUB)
{from_code} · {to_code} Their codes only DUB
{altitude} Altitude, in the posts' units 11 280 m
{speed} Ground speed, in the posts' units 820 km/h
{climb} Climbing, descending or level descending
{heading} Where it is heading, in words south-east
{arrow} The same as an arrow, north up (1.1); no word to translate ↘
{squawk} Its transponder code 7700
{why} Why it was posted: rare, military, heavy, the farthest… A rare visitor.
{area} The area you wrote yourself (SOCIAL_AREA) above Metz
{time} · {date} When, on the Pi's clock 21:42
{photo} The Planespotters page of the aircraft (when photo links are on) a link
{tracker} Its page on ADSB Exchange's map a link
{hashtags} Your hashtags #avgeek

No tag gives a distance, a direction from the antenna or a place near it: the posts never give the receiver's position away, and the templates cannot either. An unknown tag is refused when the template is saved; sudo zenithboard social template test shows what the last aircraft would have given; a template longer than Bluesky's 300 characters with plausible values is refused. Several templates rotate, in order or at random, like the styles today. Kept in /etc/zenithboard/social-templates.txt (one per line, no secret in it), editable from the console's Sharing.

What can break, and how it is kept from breaking

Risk Kept in check by
A string left in English, a translation lost when the English changes A test lists, per catalogue, what is missing or orphaned; English is always the fallback
A {placeholder} dropped or misspelt in a translation A test compares the placeholders of every entry with the English
German and Dutch about a third longer: buttons wrap, cards overflow Previews at desk and phone width in French, Russian and Japanese before each merge; long words may break (CSS hyphens)
A post over 300 characters in a longer language The posts' fitting (lists shortened, lines dropped) runs per language in the tests, at the longest plausible values
Accents and Cyrillic on the dot pictures The 5 × 7 font gains É È À Ü Ö Ä Ñ Ç Ã Õ … and the 33 + 4 Cyrillic letters, each drawn and previewed; until then they are written without accents (as the wall does today)
Chinese, Japanese, Korean on the summary cards Noto CJK, installed on demand; the card refuses to draw rather than drawing empty boxes, and says why
The English page changing by accident English pages and posts compared with today's, byte for byte
Tests that read English strings in the code Unchanged: the English stays in the code as the key

What it costs the Pi (Raspberry Pi 3B, 1 GB)

Cost
Disk About 40 KB of catalogues per language, under 0.6 MB for the twelve; +91 MB only with a Chinese, Japanese or Korean post language (Noto CJK)
Memory, the wall's server Nothing held: the catalogues are files the browser fetches once per page
Memory, the Bluesky service One language's catalogue (about 40 KB); +4.4 MB while a CJK card is drawn (measured on a desk machine, Pillow 12)
CPU Nothing measurable on the pages (the translation happens in the browser); a CJK card took 0.5 s to set its text on a desk machine, so a few seconds on the Pi 3B, once per summary
Network One catalogue (about 40 KB) per page load on the home network

The order proposed

  1. The framework and French — the console, the statistics page, graphs1090's page; the language at first sign-in and in the top bar. French first: the owner reads it, so the framework is judged on a real translation. Done, 0.47.0.
  2. The pages in every other language — German, Dutch, Italian, Spanish, Portuguese (Brazil), Polish, Finnish, Estonian, Latvian, Lithuanian, Indonesian, Russian, Ukrainian, Arabic (right to left), Japanese, Korean, Chinese (Simplified and Traditional). Done, 0.48.0. Swahili offered in 0.53.1.
  3. The posts: SOCIAL_LANG, the five post styles, the summaries and their cards, in the languages done so far; your own templates and their tags. Done, 0.51.0, for the words: every post and summary text, the alt texts, and the owner's templates (social template, a sixth style). The cards' labels wait for steps 4 and 5.
  4. The dot font's letters for the pictures: the accents of the Latin-script languages (Polish ą ę ł ś ż, the Baltic ā č ē ģ ī ķ ļ ņ š ū ž ė į ų, Finnish and Estonian ä ö õ…), then Cyrillic for Russian and Ukrainian. Arabic's pictures keep their labels in English (a 5 × 7 cell cannot join Arabic letters), like Chinese, Japanese and Korean.
  5. Japanese, Korean, Chinese (Simplified and Traditional) — Noto CJK for the cards on demand; the dot pictures' labels in English.

Each step its own release, previewed, with its review sheets. Step 1 is done (0.47.0).

What 0.47.0 measured

  • About 600 keys for French across the three pages (console 490, statistics 90, graphs1090 25): 60 KB of catalogues, one fetched per page load; nothing held by the wall's server, which only lists which languages it has.
  • The pages stay one script each: the console waits for its catalogue (a few KB, on the home network) before drawing; a missing or broken catalogue shows the page in English.

Clone this wiki locally