Fluent enough.
A language-agnostic drilling app for basic language skills — vocabulary, production, listening and grammar — scheduled by spaced repetition, running offline on your phone once a language's decks are downloaded.
Download · Courses · Decks · Contribute · Roadmap
Status: early development. The architecture, deck format, example decks, deck parser and validator are in, and the app's screens are built from the design. Progress is saved on the phone, one database per profile, and features not built yet are shown disabled, marked "Feature incoming". See docs/ROADMAP.md.
Most drilling apps are built around one language and then stretched to fit others. Fluenough inverts that: a language is just data. A deck is a text file, the card model is a superset that accommodates Latin, syllabic and logographic scripts alike, and grammar is described by pattern tables rather than hard-coded rules. Adding a language means adding files, never code.
- Each skill is scheduled separately. Seen, heard, spoken and written words, and grammar, each have their own schedule for each word: you can recognise a word long before you can produce it.
- FSRS is fitted on the phone. The scheduler learns how you forget, per language and skill, from your own answers.
- Your learning stays on the phone. There is no account, progress is saved on the phone, and nothing is sent anywhere to fit your schedule. The app goes online only to download decks, to check for updates and to send a report you choose to send. Downloading decks, like the update check, tells GitHub the phone's IP address and which files it asked for; nothing else. Once a day at most, at launch, the app asks GitHub whether its decks have changed; Settings > Deck downloads turns that off.
Taught from English: Assamese, Bengali, Gujarati, Hindi, Kannada, Marathi and
Telugu, each starting with its script decks and a script guide. Spanish is
started. Every one of these has a B1 plan in its path: the units up to B1, in
teaching order, those not yet written shown as "Coming". Telugu and Bengali
have their A1 units written in full. A course is tagged Alpha until its A1
units are written and Beta until its B1 units are. Japanese (decks/ja/) is
kept in the repository but not offered in the app for now. See
docs/ROADMAP.md.
The app comes with no decks. It downloads a language's decks from this
repository's main branch when you choose it, and you can start once its
first five are in; after that it works offline. Fixed and new decks arrive
without an app update: once a day at most it asks whether to fetch them, and
Settings > Deck downloads lists each language, with Update and Remove, and
can turn the daily check off
(ADR-0037).
A Bengali learner from English, three and a half weeks in, and the same phone in reviewer mode.
| Deck path | Today |
|---|---|
 |
 |
| Progress | Reviewing |
 |
 |
Fluenough is an Android app, built and run on Android phones for every
release. Download app-release.apk from the
latest release and
open it on the phone, allowing installs from your browser or file manager when
Android asks (sideloading).
To build it yourself, see Building below.
Each part opens in place.
Skills and drills
| Skill | Direction | Graded by |
|---|---|---|
| Seen words (recognition) | see target → recall or choose the meaning | you, or the app for a choice |
| Heard words (listening) | hear target → choose, then type, the meaning | the app |
| Spoken words (speaking) | see meaning → say target | the app, through the phone's speech recognition |
| Written words (production) | see meaning → type target | the app, with diacritic and typo tolerance |
| Grammar | understood: see a form → choose what it means; produced: prompt + slot → choose the form, then type it | the app |
In a speaking drill, after a wrong or almost-right answer, Try again lets you say the word again as often as you like, and Hear yourself plays your voice back before the model's; only your first answer is scheduled. A word's base words show on its card ("Base: जाना (jānā) · to go"), and Inspect shows, for each card, how hard it has been for you in each skill you have answered it in.
Reading passages, match pairs, multiple choice and word order come on top. Minimal-pair discrimination, for sound contrasts the learner's own language does not make, is planned; a word can already name its minimal-pair partner, which listening offers among the meanings.
All share one card model, and each word is scheduled separately in each skill: you can recognise a word long before you can produce it. A right answer also counts in part for the skills it implies, as the research finds (ADR-0034).
Spaced repetition, with a real audit trail
Scheduling is FSRS-6 (ADR-0033), fitted on the phone to each learner, per language and skill, from their own answers; nothing is sent anywhere (ADR-0035). The part that matters more is that every review is written to an append-only log — not just the current interval. That means your statistics are recomputable, and the scheduling algorithm can be replaced later without throwing away your history. Most apps store only current state and can never go back.
Text to speech
Listening drills use the operating system's own TTS — android.speech.tts
on Android, AVSpeechSynthesizer on iOS — through a pluggable TtsEngine
interface. This adds nothing to the download size, needs no network, and
covers far more languages than any bundled model could. Users install voices
through system settings, and Settings > Voices picks which of the phone's
voices speaks each language.
A neural backend (Kokoro) is a candidate for a later release; see ADR-0002 for why it is not in v1.
Scripts are first-class
Where a language uses an unfamiliar writing system, learning the script and its
pronunciation is the first task, not a preliminary to it. Fluenough treats
script decks as ordinary decks — decks/hi/hi-en-script-vowels.yaml is a
worked example. The courses taught from English (Assamese, Bengali, Gujarati,
Hindi, Kannada, Marathi and Telugu) each start with their script decks and a
script guide; Spanish is started. See docs/ROADMAP.md.
Japanese (decks/ja/) is kept in the repository but not offered in the app
for now.
How words are written in Latin letters: ISO 15919
Every word in an Indian language has a reading, its Latin letters, shown on the card. The readings are based on the letters of ISO 15919, the international standard for writing Indian scripts in Latin letters, and they depart from it wherever a learner is better served (ADR-0025). They are not a strict ISO 15919 transliteration: the list of departures follows the table of letters.
ISO 15919 was chosen because its marks show what plain letters cannot: పాలు, milk, is pālu and పలు, many, is palu; పాట, song, is pāṭa and పాత, old, is pāta. One standard covers all the Indian scripts, so a learner of a second Indian language meets the same letters again.
ISO 15919 writes letters, so कितना is kitanā letter for letter. The decks write each word as it is said, in ISO 15919's letters: kitnā, without the a that Hindi does not say. Where a language says two letters alike, the reading writes the sound: Bengali ঈ is i, as ই is.
| Letters | Sound | Example |
|---|---|---|
| a | the short vowel every consonant carries: u in but in Hindi, a in Telugu | कल kal |
| ā ī ū | long a, i, u: held about twice as long | పాలు pālu, नहीं nahīm̐ |
| e ē, o ō | short and long e and o, where a language has both (Telugu, Kannada) | నేను nēnu |
| ē ō | the long e and o of Hindi, which has no short ones | मेरा mērā |
| ai au | the vowels of ऐ and औ | है hai |
| ô | Bengali's and Assamese's own vowel, as in law | কথা kôthā |
| ê | the vowel of cat, in Bengali, and in words from English | ব্যাগ bêg |
| ṭ ṭh ḍ ḍh ṇ | tongue curled back to the roof of the mouth (retroflex) | పాట pāṭa, अंडा aṇḍā |
| t th d dh n | tongue on the back of the upper teeth (dental) | పాత pāta |
| kh gh ch jh th dh ph bh | an h after a letter is a puff of breath, never one sound: th is not thin, ph is not phone | खाना khānā |
| c ch | ch of church; ch with a puff of breath | चाय cāy, छह chah |
| ś ṣ | sh; ṣ with the tongue curled back | शादी śādī |
| ṅ ñ | ng of sing; ny of canyon | বাংলা bāṅlā |
| ḷ | an l with the tongue curled back | ನಾಳೆ nāḷe |
| ṛ ṛh | a flap with the tongue curled back | लड़का laṛkā |
| ṁ | the anusvara where it nasalises the vowel before y, r, l, v, s or h | संसार saṁsār |
| m̐ | the vowel before is said through the nose | हाँ hām̐ |
| k͟h q ġ z f | sounds from Persian, Arabic and English, written with a dot (nukta) | ख़त्म k͟hatm |
| x | Assamese's own sound, as in Scottish loch; ISO 15919 has no letter for it | অসম ôxôm |
Where the readings depart from ISO 15919
Each row is a deliberate departure, made because the learner is better served
by the sound than by the spelling. "ISO 15919 writes" is the letter-for-letter
reading of the script, as the tables in tools/transcribe.py give it. The
last column is the file that shows it: a card id, or the language's
decks/<code>/<code>-romanisation.yaml, which says the same in a paragraph.
| What | ISO 15919 writes | Fluenough writes | Why | Shown by |
|---|---|---|---|---|
| Inherent a, where it is not said (Hindi, Marathi, Gujarati, Bengali, Assamese) | कितना kitanā | kitnā | Letter for letter suggests a syllable nobody says | hi-0163; hi-romanisation.yaml |
| Inherent a in Bengali and Assamese | কত kata | kôto: ô where it is said [ɔ], o where it is said [o] | The vowel is not [a]. ô is ISO 15919's letter for ऑ, reused | bn-0417, as-0377; bn-romanisation.yaml, as-romanisation.yaml |
| য-ফলা, ব-ফলা and ম-ফলা (Bengali, Assamese); ê for [æ] (Bengali) | স্যার syāra, আত্মীয় ātmīẏa | sêr, āttiyo | The mark holds the consonant or changes the vowel; it does not say y, v or m | bn-0187, bn-0443; bn-romanisation.yaml |
| Conjuncts said otherwise than their letters | ज्ञ jña, क्ष kṣa | ज्ञ: Hindi gya, Marathi dnya, Gujarati gna, Bengali gg; क्ष: Hindi kśa, Gujarati kśa, Bengali and Assamese kkh (Marathi क्ष stays kṣa, as ISO 15919 writes it) | The letters are not how the conjunct sounds | hi-0083, hi-0085, mr-0089, gu-0085, bn-0081, as-0085 |
| Anusvara ं and the nasal vowel | always ṁ: अंडा aṁḍā, नहीं nahīṁ | the nasal said, aṇḍā; m̐ for a nasal vowel, nahīm̐; ṁ only before y, r, l, v, a sibilant or h, kiṁvā; Bengali and Assamese ং is ṅ, bāṅlādeśi. The same in Gujarati, Telugu and Kannada: dukāṇam, tiṅgaḷu, navembar (ISO 15919: ṁ) | Anusvara is whichever nasal comes next | hi-0258, hi-0156, mr-0577, bn-0422, te-0246, kn-0451, kn-0464; hi-romanisation.yaml, bn-romanisation.yaml, te-romanisation.yaml, kn-romanisation.yaml, gu-romanisation.yaml, as-romanisation.yaml |
| Telugu's half nasal ఁ | teravam̐baḍalēdu | teravabaḍalēdu, left out | It is no longer said | te-en-reading-diddubatu.yaml; te-romanisation.yaml |
| Long ī ū, where the language does not tell them from i u (Marathi, Gujarati, Bengali, Assamese) | मी mī, কী kī | mi, ki | A mark would suggest a contrast the language does not make. Bengali ঈ is i | mr-0206, bn-0193, gu-0310 |
| ए and ओ in those four languages | ē ō: आहे āhē, মোড় mōṛa | e, o: āhe, moṛ | No short e or o to tell them from | mr-0205, bn-0314 |
| Bengali ঐ ঔ; Assamese ঐ ঔ and the vowel sign ো | ai au; Assamese মোৰ mōra | Bengali oi, ou; Assamese ôi, ôu, and u for ো in words, said [ʊ] (mur, muk). The letter ও alone still reads o | They are said so | bn-0775, bn-0011, as-0188, as-0167, as-2064 |
| Bengali স and ষ (শ is ś in both) | s and ṣ: সে se, পরিষ্কার pariṣkāra | ś where said sh: śe, poriśkār; s where said s (strī) | Said alike, as sh or as s, by the word | bn-0126, bn-0683; bn-romanisation.yaml |
| Assamese স শ ষ | s ś ṣ: দেশ dēśa | x: dex | One sound, as in Scottish loch; ISO 15919 has no letter for it | as-0173 |
| Other Assamese letters said alike | ছাৰ chāra, জানুৱাৰী jānuvārī, ভাইটি bhāiṭi, ভণ্টি bhaṇṭi | sār, zānuwāri, bhāiti, bhonti: চ ছ as s; জ ঝ য as z; ৱ as w; ṭ ṭh ḍ ḍh as t th d dh; ণ as n | Assamese says them alike | as-0162, as-0393, as-0155, as-0156 |
| Other Bengali letters said alike | কারণ kāraṇa; য y, ঢ় ṛh, য় ẏ | kāron; j for য, ṛ for ড় and ঢ়, y for য় | Bengali says them alike | bn-0365, bn-0739; bn-romanisation.yaml |
| ष in Hindi and Gujarati | भाषा bhāṣā, શિક્ષક śikṣaka | bhāśā, śikśak | Said as श | hi-0197, gu-0322 |
| ऋ | r̥: कृपया kr̥payā | Hindi kri, Marathi, Gujarati, Kannada and Telugu ru, Bengali and Assamese ri | How it is said, in letters every keyboard and font has | hi-0616, mr-0157, te-0007, as-0007 |
| Kannada ೞ; Kannada ಱ and Telugu ఱ | ಮೞೆ maḻe; ṟ | maḷe; r | No longer said differently from ಳ and ర | kn-0580; kn-romanisation.yaml, te-romanisation.yaml |
| ಫ in words from English | ಕಾಫಿ kāphi | kāfi | Said f | kn-0494 |
Each language's file in decks/<code>/<code>-romanisation.yaml says how its
readings are written, in a paragraph. Each card also records how the word is
said in the IPA, the International Phonetic Alphabet: /paːlu/. The app
does not show it yet.
A typed answer needs none of the marks. palu, paalu and pālu are all
right for పాలు, as kitna and kitnaa are for कितना: answers are compared
after the marks are taken away, and each romanisation file lists the spellings
people type for one sound. To write a reading and IPA for a new card, run
python3 tools/transcribe.py <code> "<word>" <how it is typed>.
Decks are YAML files under decks/, versioned in git like any other
source. They are diffable, reviewable, and contributed as pull requests. The
app downloads them from main, by decks/index.json, so a merged deck
reaches learners without a release.
A deck, and how to validate it
schema: 1
id: es-en-core-100
name: Spanish Core 100
language: { code: es, iso639_3: spa, name: Spanish, script: latin, tts: es-ES }
native: { code: en, iso639_3: eng, name: English }
license: CC-BY-SA-4.0
cards:
- id: es-0001
target: la casa
native: the house
tags: [noun, home]The full specification is in docs/DECK-FORMAT.md. Validate any deck before opening a pull request:
python3 tools/deck_index.py # write decks/index.json again
python3 tools/validate_decks.py decks/The validator needs only Python 3.11+ and PyYAML — no Flutter toolchain. It
fails while decks/index.json is out of date.
The Telugu, Marathi, Kannada, Gujarati, Assamese and Spanish decks have not been checked by a speaker
A deck no speaker has checked yet is tagged unreviewed and says so: some
Hindi and Bengali decks, and all the Telugu, Marathi, Kannada, Gujarati,
Assamese and Spanish ones. The Telugu decks (decks/te/) were written from published
sources; the Marathi (decks/mr/), Kannada (decks/kn/), Gujarati
(decks/gu/), Assamese (decks/as/) and Spanish (decks/es/) decks are a
first attempt written for Fluenough. No speaker has checked any of them
yet. Each one says so in its description and on its screen in the app. If you
speak one of these languages, turn on Review decks in Settings: the phone makes
your rater code, and on each unit you mark cards right or suggest a change, then
send the reviews in one mail from your own mail app
(guide for reviewers). Offensive words are not in an
ordinary review: each language has an Offensive words review of its own, which
explains why they are in the app, asks whether you are 18 or over every time,
and has you rate each word and check the words that sound like one. Each card
records the rater codes of the speakers who checked it, and a deck turns
reviewed once every card is checked. Or
report mistakes or send a
fix; a review by a speaker is the most useful contribution these decks could
get.
Sources
Most decks are written for Fluenough. Where a deck's text comes from
elsewhere, the deck names the source in its source field, and the app shows
it: under each passage, on the deck's page, and in Settings, under Sources.
The Bengali decks quote two books in the public domain, letter for letter:
- Sahaj Path, part 1, by Rabindranath Tagore (1930), in the public domain: passages for reading.
- Abol Tabol by Sukumar Ray (1923), in the public domain: older spellings in the spelling deck, and passages for reading to come.
The Hindi reading deck quotes a story in the public domain, letter for letter:
- Panch Parmeshwar by Premchand, as printed in Prem-Dwadashi (1926): passages for reading, checked against the page images.
The Telugu reading deck quotes a story in the public domain, letter for letter:
- Diddubatu by Gurajada Apparao (1910), as printed in Gurujadalu (2012): passages for reading, checked against the page images.
A new source gets a line here, and its decks name it in source.
Building from source
You need the Flutter SDK (3.47+) and, for Android, a JDK and the Android SDK. The platform folders are not committed; generate them on first checkout:
flutter create . --org app --project-name fluenough --platforms=android,ios
rm -f test/widget_test.dart # generated boilerplate; references MyApp, not ours
flutter pub get
dart run build_runner build
flutter runSee docs/DEVELOPMENT.md for toolchain setup, including notes for immutable Linux distributions and a warning about Waydroid and TTS.
Decks are the most useful contribution and need no Flutter toolchain — just a text editor and Python. Code contributions are welcome too; the roadmap says what is wanted next.
- CONTRIBUTING.md — how to contribute, deck or code
- AGENTS.md — how to work in this repository without colliding with the other people doing so. Required reading before your first code change, whether or not you use a coding agent.
- docs/adr/ — why things are the way they are
Several people build this in parallel, most of them with agents. Claim an issue before you start, say which files you expect to touch, and keep pull requests to one concern each.
GPL-3.0, with an App Store Distribution Exception. See LICENSE and LICENSE-EXCEPTION.md.
The licence in detail
Fluenough is free software and every fork must stay free software. The exception exists solely so the app can be distributed through stores whose terms would otherwise conflict with the GPL; it grants no permission to close the source. See LICENSE and LICENSE-EXCEPTION.md, and ADR-0003 for the reasoning.
Deck content carries its own licence, declared per file.
The pictures on cards are from Noto Emoji, by Google, under the Apache License 2.0; the licence ships with them in assets/pictures/LICENSE.