Speedready

Pacer and RSVP reader for epub/txt, built for language learners. GTK4 / libadwaita, one Python file.

• Pacer: the whole text on screen, a highlight sweeps through at your pace, read words dim behind it
• RSVP: one word (or chunk) flashed at a time with a red pivot letter
• First-run language setup; the interface and translation target follow the learner's language (English and Brazilian Portuguese UI)
• Select text in the pacer and send the phrase to DeepL or Google Translate. In beginner mode, long-press the first word and tap the last one
• Click a word to continue from there. Double-click it for a dictionary popup (lemmatized; en.wiktionary glosses plus the Duden meanings for German, browser button for the full page); close it and the flow resumes from that word
• Right-click a word you don't know: it gets a subtle underline everywhere it occurs and the pacer lingers on it. Your list lives in unknown.txt
• P reads the current sentence aloud (piper TTS, offline, voice downloaded on first use); the popup has a pronounce button
• A read-along: piper reads sentence by sentence and the highlight follows the voice. WPM sets the speech speed
• Chapter list from the epub's TOC (works around broken anchors by finding the headings in the text)
• Every lookup lands in ~/.config/speedready/vocab.tsv with its sentence, importable into Anki as-is
• Remembers your position, WPM, chunk size, reading mode, beginner mode, and read-along setting per book; progress is saved atomically while you read
• Bookmarks are stored per book (B toggles one, Shift+B opens the list)
• Library (L): every book you have opened, newest first, with how far through it you are. Add several at once from the file picker, or drag epub and txt files straight onto the window. Titles come from the epub's own metadata, so the list reads like a shelf and not like a download folder
• Everything is configurable in the settings dialog (fonts, colours, pauses, chunk size, dictionaries)

Install (Linux, GTK4 + libadwaita + python-gobject)

./install.sh        # venv with simplemma + piper-tts, launcher + icon
./run.sh book.epub  # or open "Speedready" from your app launcher

Keys

key	action
space	play / pause
← →	±10 words
PageUp / PageDown	±one page
↑ ↓	speed ±5 (shift: ±25)
[ ]	words per step
M	pacer / rsvp
R	replay sentence
D	define current word
click / double-click / right-click	continue from there / dictionary popup / mark unknown
P	speak from here to the end of the sentence, again to stop
A	read-along
B / Shift+B	toggle bookmark / open bookmarks
C	chapters
L	library
F11, S, O	fullscreen, settings, open

Dictionary packs

Beginner mode glosses each word from an offline pack, gloss-<src>-<tgt>.sqlite, which the app downloads once per language pair on first use. The packs are built from English Wiktionary and each language's own Wiktionary edition (kaikki.org) and hermitdave/FrequencyWords. See DATA-LICENSES.md for upstream licence and attribution details.

python tools/fetch_sources.py            # ~2 GB of extracts -> ~/.cache/speedready/build
python tools/release_packs.py --jobs 6   # build every pair, measure, keep the good ones
python tools/release_packs.py --upload   # gh release upload to the "packs" release

A pack is only published if it glosses at least half of the 2000 commonest words in its source language (--min-coverage). Pairs below the floor are dropped rather than shipped, because a pack that exists but glosses nothing would have the app report the dictionary ready over a page of blank glosses. An unpublished pair is reported as such in the app. MUSE-free pack build, 2026-09-25: all 90 language pairs passed the 50% threshold. Median coverage was 92.1%; best was sv->en at 96.7%, worst ru->nl at 54.8%. Total size was 4.66 GB. These figures come from tools/release_packs.py measured against each pair's top 2,000 source words.

Content words are reliable across every pair. The commonest grammar words often are not: Wiktionary picks the wrong homograph for them (sv->de gives är as "Gut, Ware", nl->pt gives is as "seu"). GLOSS_OVERRIDES in speedready.py fixes these by hand for de->pt only. Curating that list once for de->en and pivoting it would fix every de->X pair at once, since knowing that wollen is "want" and not "woollen" does not depend on the target language.

German noun genders

In German books every noun is colored by gender: blue for der, red for die, green for das (Settings → Deutsch). gender.py holds the logic, with no GTK code so a browser extension can port it line for line. It takes the gender from the article in front (die stille See is the sea, der See the lake), splits unknown compounds (Hahnenfeder → Feder), colors plurals by their singular, and leaves pronouns, names, plural-only nouns and ambiguous sentence-initial words alone. web/gender.js is the same logic for the web reader, and web/textclean.js is the book cleanup; tests/test_web_parity.py fails if the Python and JavaScript versions ever disagree (PYTHONPATH=. python tests/test_web_parity.py *.epub checks real books). The data, web/genders-de.json.gz (0.6 MB, noun form → genders), comes from the kaikki.org English Wiktionary extract:

python tools/build_genders.py ~/.cache/speedready/build/en-German.jsonl.gz
python gender.py   # self-check

Web version (any phone or browser)

web/ is a self-contained reader for learners, and it is also what the Android and iOS apps load: a book in any of 10 languages with a translation in your own language above every word you have not marked as known yet. Tap a translation to mark the word learned and it disappears everywhere; tap a word for its dictionary entry. Books you add stay in a library (the apps can also scan a folder for books and let you pick which to add). The interface is in all 10 languages. It ships with a public-domain German book, speaks through the device's own voice, and works offline once loaded (installable to the home screen). It is a separate implementation from the desktop reader; they share the data pipeline (tools/export_web.py, coreglosses.py), not the reading code.

core/ holds hand-checked glosses for each language's ~300 most frequent words, which a context-free dictionary gets wrong most often. python tools/export_web.py --patch <dict files> writes them into exported dictionaries; pairs without English go through English.

Read-along drives the highlight from its own estimated clock rather than from onboundary, which iOS reports unreliably, and re-syncs whenever a boundary event does arrive. Devices with no speech voices installed say so instead of failing silently.

Build its data from a desktop pack:

python tools/export_web.py de pt      # -> web/dict-de-pt.json (8.5 MB, ~1.7 MB gzipped)

The shipped language pair is named in web/bundle.json and nowhere else; index.html and sw.js both read it, and the reader takes its source language from the dictionary's own src field. So changing which pair the website and the Android APK carry is that one command plus a starter book, with no code edit. A test enforces it: neither file may name a dictionary or a language again.

One APK still ships one pair, because the dictionary is bundled and the app asks for no permissions at all. Shipping more than one means either a bigger APK or network access, which is a trade worth making deliberately.

iPhone and iPad app

ios/ is a native shell around web/, the same way android/ is: the page does the reading, and ios/Speedready/App.swift serves it offline from the app bundle and adds dialogs, the Anki export via the share sheet, links out to Safari, and speech that plays with the silent switch on. It needs iOS 16.4+ (the offline dictionary unpacks with DecompressionStream).

open ios/Speedready.xcodeproj                      # run in the simulator from Xcode
ASC_KEY_PATH=… ASC_KEY_ID=… ASC_ISSUER=… ios/release.sh   # signed .ipa for App Store Connect, from Terminal on the Mac

Create the app record (bundle id io.github.cyoren.speedready) in App Store Connect before the first upload.