Anki-like flashcard app that teaches basic maths to children 😉.
No Python, no terminal, no admin password. Grab the file for your system from the latest release:
| System | File | First launch |
|---|---|---|
| macOS (Apple Silicon) | Almathus-*-macos-arm64.dmg |
The build is ad-hoc signed, not notarised: right-click the app → Open the first time, or macOS calls it damaged |
| Windows 10/11 | Almathus-*-windows-AMD64.exe |
It is unsigned, so SmartScreen shows a warning: More info → Run anyway |
| Linux x86_64 | Almathus-*-linux-x86_64.AppImage |
chmod +x and run it; the .tar.gz is there if AppImage is not your thing |
Each release also carries SHA256SUMS, the GPL source of the voice engine
(almathus-voice-<version>.tar.gz), and inside every bundle a licenses/
folder with the full texts.
The exercises are the same ones speech and language therapists give to help children build automatic recall of basic operations.
The game is aimed at children aged 6 to 8, and for now covers addition, subtraction, multiplication and division. Harder levels will be added later.
Each game is a deck of two-number cards: the child gets one card at a time, answers it, and the next one comes up. A stopwatch times every answer and the whole game. The faster the answer, the better.
Parents can organize daily sessions with their child, follow progress made overtime with detailed reporting.
The interesting part of this MVP is the voice loop: TTS reads the question aloud, STT takes the answer, and the game validates it. The keyboard and the on-screen keypad also work, but the aim was to keep the child's hands off the pad to solely focus on the solving part without distractions.
Built with Python 3.14 and Tkinter, and no third-party packages: the base game
runs on the standard library alone. The voice features are optional
(--extra voice). Speech recognition runs on Vosk models, one per language and
downloaded on first use; speech out runs on Piper, in a separate process. uv
installs a Python build that bundles Tk 9 on macOS, Windows and Linux. A system
Python needs its own Tk package.
Two theme skins: light and dark modes.
Light A.K.A "Princess" mode
Dark A.K.A "Prince" Mode (the default palette)
uv run almathus
# or
uv run python main.py
# or, without uv (needs Python 3.14 and its Tk):
python3.14 main.py- Set up the game. Choose how many cards to put in the deck (20 by default),
then pick one operation: addition, subtraction, multiplication or division.
The number row runs from 0 to 10 and starts with all of them on. Division
ignores 0, because there is no division by zero. Skip easy cards leaves out
subtraction above 10 that needs no borrow:
19 − 9is9 − 9with a ten left. It is on by default and remembered once switched off. - Wait for the countdown. 3, 2, 1, go.
- Answer each card. The operation is shown in big numbers (
8 − 2). Tap the numbers on the on-screen keypad, type them and press Enter, or just say the answer out loud. - Expect a second pass. Every card missed on the first try comes back once at the end. 🔁
- Read the recap. Score, total and average time, and the five slowest operations, the ones worth practising tomorrow.
To abandon a game, tap ← Menu at the top left of a card or the speak-clearly screen. It asks before leaving a card and records nothing. That is useful when the wrong operation or deck size was picked.
My progress, accessible from the top left of the settings screen or on the recap after a game, shows:
- Average session time. On the left-hand axis, one point per day on a real calendar timeline.
- Grades. On the right-hand axis, out of the deck the child plays most. A day with no session leaves a gap in that line.
- A date filter. All months, or any one month from the list.
- Every session. Newest first: its grade, its average time per card, and the five cards it fought with.
- A PDF. Same chart, same table, in the colors on your screen.
A game is timed twice over: the current card (bottom-left) and the session as a
whole (bottom-right). When the game ends, the session is saved twice, in
Documents/Almathus/ or wherever ALMATHUS_RESULTS_DIR points:
| File | What's in it |
|---|---|
results.csv |
one line per session. This is the one you can edit by hand, and the one the progress screen reads |
stats/<timestamp>-<mode>.json |
everything about every card: time, attempts, whether it was right first time |
The app's database: one line per session, plus the columns the progress screen reads. Every finished session lands at the bottom, and whatever you edit by hand is what the app believes next time. It only ever appends. The one thing it will touch is the header, which gains a column when your file is missing one the app knows about.
date,family,time,cards,grade,average_time_per_card,difficulty
2026-06-16,-1,0:31,,,,
2026-07-28,-2 and -3,1:35.4,20,18,4.8,"17 - 4, 13 - 4, 9 - 4, 6 - 4, 20 - 4"| Header | Meaning |
|---|---|
date |
day the session was played |
family |
which family was drilled: -3, -2 and -3, x7 |
time |
total time of the session |
cards |
how many cards were asked |
grade |
cards right on the first try |
average_time_per_card |
time / cards |
difficulty |
the five cards that session fought with |
The app writes the header in whatever language it is running in, so a French
install writes the same seven columns in French, and both are read in any
column order. Matching ignores case, accents, spaces and underscores (Temps,
Average time per card and temps_moyen all work), and a few synonyms are
accepted, listed in ALIASES (results.py). Columns the app does not know are
left untouched when it appends.
| Column | What you can write |
|---|---|
| date | 2026-06-16, 16/06/2026, 16/06/26, 16-06-2026, 16.06.2026, 2026/06/16, or short 16/06 · 16-06 · 16.06. Empty repeats the row above. |
| family | free text. The app writes the operator sign then the numbers switched on: -3, -2 and -3, x7, +1–10. ASCII signs (+ - x /) so the column stays easy to type. |
| time | 1:23 · 1:23.4 · 1'23 · 1'23"4 · 83" · 83 · 83.4 · 1,23. A comma followed by exactly two digits below 60 is the minute separator (1,29 = 1 min 29 s); otherwise it is a decimal point (32,60 = 32.6 s). Use a dot for a decimal. |
| cards | an integer, or empty when unknown. |
| grade | cards right on the first try, as a bare number (15); the denominator is the cards cell beside it. Typing 15/20 keeps only the numerator. |
| average time per card | plain seconds. Empty means it is computed as time / cards. |
| difficulty | the hardest cards, hardest first: 18 - 4, 11 - 4, 13 - 4. The next session reads this cell back, so writing a card in is a way to say drill this tomorrow. |
Rows that can't be read are skipped, never rewritten. A fresh install starts with the header line and no rows.
- What makes a card difficult. A wrong answer, or a slow one. The session
ranks its cards, wrong first, then the slowest. The top five go into the
difficultycolumn, so you can focus on the operations the child struggled with most. - Both sittings count. A card missed on the first pass comes back at the end, and the two times are added, so a quick fumble keeps the time it really cost.
- Every listed card comes back twice, the copies kept apart in the shuffle. Back to back, the second would only test the memory of the first.
- Only what the round can ask. It reads the last day played, all of that day's sessions together, and deals the same operation from a number that is still switched on. Switch from −4 to ×2 and the −4 cards wait.
- Never more than half the deck. Five carried-over cards need a deck of twenty.
- The average here is the whole session. The recap averages thinking time per card; this column divides the whole session, pauses and retry round included, by the card count, so it matches the columns beside it.
| What | Where |
|---|---|
| Number ranges | RANGES in main.py: addition and subtraction use operands 0–20 (never negative); multiplication and division use factors 0–10 (division always divides evenly). The number row is 0–10 and starts fully on; division skips 0 |
| Easy cards | Skip easy cards on the settings screen: leaves out subtraction above 10 that needs no borrow (19 − 9, 14 − 4), so only cards that cross a ten remain. On by default; ignored when it would leave fewer than five different facts |
| Number words | generated from num2words, except Chinese, Hindi, Gujarati, Georgian, Kyrgyz and Uzbek, written out in _TABLE_NUMBERS (voice.py) |
| Language | chosen in the app (App language on the settings screen, or the chip in the top-right corner) and remembered in prefs.json; DEFAULT_LANG in main.py only decides what an install starts on |
| Remembered choices | prefs.json: language, dictation, theme, number selection, easy-card switch |
| Results folder | ALMATHUS_RESULTS_DIR |
voice.py: speech out and speech in, each a no-op when its dependencies are
missing. The keypad always works.
Piper, one model per language (63–128 MB), downloaded at setup. The game never
imports it: it runs sidecar/piper_sidecar.py and exchanges text and PCM over
stdin/stdout, which keeps the GPL engine out of the MIT app. Source runs need
--extra sidecar; packaged builds ship a frozen helper.
- The voice is a file, so a card sounds the same on a Mac and on a PC: nothing to enable, no language pack, nothing online.
- No Piper model for Esperanto, Gujarati, Kyrgyz, Tajik or Uzbek, and none until a
download finishes: the OS voice takes over (
say,espeak-ng,pyttsx3), picked from that language's adult voices, women first. ALMATHUS_TTS_VOICEpicks another voice.ALMATHUS_TTS_BACKENDforces a backend:say,espeak,pyttsx3orpiper(autoby default).
French numbers are read the way a teacher says them. six, huit and dix
keep their final consonant before an operator, and plus is /plys/ rather than
the /ply/ of conversation. The text is respelled before it is voiced (dix →
disse, huit → huite). French is the one language of the 23 measured where a
number loses a consonant in context.
Each card is read aloud before the child answers, on by default. The mic opens when the reading ends and neither clock runs during it, so the log measures thinking rather than speech. With dictation off, the mic and the clocks start with the card.
Runs continuously beside the keypad. A spoken number is filled in and submitted; the keypad stays usable throughout.
┌── the room, measured ──┐
▼ ▼
mic audio ─▶ [VAD record] ─▶ [is it a voice?] ─▶ [subtract the room] ─▶ Vosk ─▶ [numbers only?] ─▶ submit
(voice.py) level + pitch spectral subtraction per-language strict rule
Every stage can only say not an answer, and the card stays up:
- The room is measured at startup and kept current between answers. Thresholds are a factor over that measurement, so a bedroom and a terrace with a road both work, and a voice two meters back still clears the bar.
- Recording runs from the first speech to ~0.35s of silence, with a 6s cap: no button, no deadline on thinking.
- Voicing (the pitch of vocal folds) keeps chairs and coughs out of the
recognizer. Level and signal-to-noise against the room are the whisper gate, so
a murmur never scores a card; a voice too quiet gets "🔊 Plus fort !".
uv run python mic_check.pyprints what your own mic measures. - The room's spectrum is subtracted from the capture, gently (18 dB at most), and skipped when the room is already quiet.
- Vosk recognizes it, vocabulary limited to number words (~100 ms). One small model per language.
- Numbers only: every word must be a number, the all-numbers hypothesis must
come within
ALMATHUS_VOICE_ALT_GAPof the best, and a second un-constrained decode can veto it. One stray word and the utterance is refused. - A number that gets through is the answer, as if ✓ had been pressed, right or wrong.
First run downloads that language's model into a per-user folder:
~/Library/Application Support/Almathus on macOS, %LOCALAPPDATA%\Almathus on
Windows, ~/.local/share/almathus on Linux. ALMATHUS_VOSK_MODEL and
ALMATHUS_MODELS_DIR override the paths.
| Language | English | Model | Size |
|---|---|---|---|
| العربية | Arabic | vosk-model-small-ar-0.3 |
104 MB |
| العربية (تونس) | Arabic (Tunisian) | vosk-model-small-ar-tn-0.1-linto |
166 MB |
| Català | Catalan | vosk-model-small-ca-0.4 |
43 MB |
| 中文 | Chinese | vosk-model-small-cn-0.22 |
44 MB |
| Čeština | Czech | vosk-model-small-cs-0.4-rhasspy |
46 MB |
| Nederlands | Dutch | vosk-model-small-nl-0.22 |
40 MB |
| Esperanto | Esperanto | vosk-model-small-eo-0.42 |
44 MB |
| فارسی | Farsi | vosk-model-small-fa-0.42 |
53 MB |
| Français | French | vosk-model-small-fr-0.22 |
42 MB |
| ქართული | Georgian | vosk-model-small-ka-0.42 |
46 MB |
| Deutsch | German | vosk-model-small-de-0.15 |
46 MB |
| ગુજરાતી | Gujarati | vosk-model-small-gu-0.42 |
108 MB |
| हिन्दी | Hindi | vosk-model-small-hi-0.22 |
44 MB |
| English (India) | Indian English | vosk-model-small-en-in-0.4 |
38 MB |
| Italiano | Italian | vosk-model-small-it-0.22 |
50 MB |
| 日本語 | Japanese | vosk-model-small-ja-0.22 |
50 MB |
| Қазақша | Kazakh | vosk-model-small-kz-0.42 |
60 MB |
| 한국어 | Korean | vosk-model-small-ko-0.22 |
87 MB |
| Кыргызча | Kyrgyz | vosk-model-small-ky-0.42 |
51 MB |
| Polski | Polish | vosk-model-small-pl-0.22 |
53 MB |
| Português | Portuguese | vosk-model-small-pt-0.3 |
32 MB |
| Русский | Russian | vosk-model-small-ru-0.22 |
46 MB |
| Español | Spanish | vosk-model-small-es-0.42 |
40 MB |
| Svenska | Swedish | vosk-model-small-sv-rhasspy-0.15 |
304 MB |
| Тоҷикӣ | Tajik | vosk-model-small-tg-0.22 |
52 MB |
| తెలుగు | Telugu | vosk-model-small-te-0.42 |
61 MB |
| Türkçe | Turkish | vosk-model-small-tr-0.3 |
37 MB |
| English (UK) | UK English | vosk-model-small-en-gb-0.15 |
43 MB |
| Українська | Ukrainian | vosk-model-small-uk-v3-small |
144 MB |
| English (US) | US English | vosk-model-small-en-us-0.15 |
41 MB |
| Oʻzbekcha | Uzbek | vosk-model-small-uz-0.22 |
51 MB |
| Tiếng Việt | Vietnamese | vosk-model-small-vn-0.4 |
34 MB |
All 32 answer by voice. Thirty take their number words from num2words; the six
it cannot spell (Chinese, Hindi, Gujarati, Georgian, Kyrgyz, Uzbek) are written out
in voice.py. Five have no Piper model (Esperanto, Gujarati, Kyrgyz, Tajik,
Uzbek) and speak through the OS voice.
Tested on a real microphone: French and English (US, UK, India). The other 27
are wired the same way and untested. Say which language you picked, what you said
and what happened if a number is misheard; ALMATHUS_VOICE_DEBUG=1 prints the
recognizer's output to include.
Optional, with measured defaults:
| Variable | Default | What it does |
|---|---|---|
ALMATHUS_VOICE_ALT_GAP |
0.5 |
how far behind the best hypothesis an all-numbers one may score |
ALMATHUS_VOICE_VETO_CONF |
0.9 |
how sure the un-constrained model must be to overrule the number grammar |
ALMATHUS_VOICE_MIN_SNR |
3.0 |
how far a voice must stand above the room |
ALMATHUS_VOICE_MIN_VOICED |
0.25 |
fraction of frames that must carry vocal-fold pitch |
ALMATHUS_VOICE_LEVEL_OVER_ROOM |
3.0 |
how far over the measured room "out loud" is; raise it if noise scores cards |
ALMATHUS_VOICE_MIN_LEVEL_FLOOR |
45 |
never ask for less than this, so a silent room cannot leave the gate at nothing |
ALMATHUS_VOICE_MIN_LEVEL |
(unset) | a hard absolute gate; set it and nothing is measured |
ALMATHUS_VOICE_CALIB_SECS |
1.5 |
how long the startup listen to the room lasts |
ALMATHUS_VOICE_DENOISE |
auto |
1 always subtract the room, 0 never, auto when there is noise |
ALMATHUS_VOICE_DENOISE_ALPHA |
1.5 |
how hard to subtract it |
voice.py also reads ALMATHUS_VOICE_NACF, ..._ONSET_*, ..._MIN_LEVEL_BLIND
and ..._DENOISE_FLOOR for rooms that defeat the above; leave those alone.
uv sync --extra voice --extra sidecar --group dev
# voice : sounddevice + numpy + num2words + vosk + webrtcvad (the mic)
# sidecar : piper-tts, the GPL engine the separate voice process runs
# keep --extra voice on every sync: without it uv prunes the audio libs and
# the mic goes deadFirst run then asks for a language and downloads its model and voice. macOS needs microphone permission for your terminal app, under System Settings → Privacy & Security → Microphone.
ALMATHUS_VOICE_DEBUG is an environment variable; the app takes no arguments.
Set it and voice.py traces the pipeline to stderr, prefixed [voice]:
ALMATHUS_VOICE_DEBUG=1 uv run almathusIt reports capture setup, why a recording ended, the raw Vosk alternatives, why an
utterance was refused and any model load failure. Each capture is also copied to
almathus_capture.wav in the system temp folder.
uv run python test_voice.py # what counts as a spoken answer (no mic)
uv run python test_room.py # the acoustic gates, against a synthesized room
uv run python test_game_flow.py # a card leaves the screen only when answeredtest_room.py checks that the bar moves with the noise and that a voice two
meters out in a quiet room is accepted. test_game_flow.py opens a Tk window and
stubs the mic: refused utterances may not move the card, a recognized number
scores at once, and both dictation modes open a card correctly.
The repository is split by component, and the split is the point:
| Part | What it is | Licence |
|---|---|---|
The game's code — main.py, voice.py, i18n.py, results.py, … |
the Python game | MIT (LICENSE) |
sidecar/ |
separate program that owns the voice engine | GPL-3.0 (sidecar/LICENSE) |
piper-tts, espeak-ng (inside the sidecar) |
neural TTS and its phonemizer | GPL-3.0 |
Voice models (.onnx) |
downloaded at setup, not in the repo | per-model; mostly MIT / CC BY, some CC BY-NC |
assets/sounds/*.mp3 |
the four game sounds | CC0 and CC BY 3.0/4.0 (CREDITS) |
assets/flags/*.svg |
flag-icons 7.5.0 | MIT (README) |
assets/icon.svg, assets/icon.icns |
the app icon | VistaCreate stock artwork, not MIT |
num2words |
number words | LGPL-2.1 (commercial use allowed) |
vosk |
speech recognition | Apache-2.0 |
The game never imports Piper. That boundary is the reason the voice is a
separate program: the app starts sidecar/piper_sidecar.py as another process
and talks to it over stdin/stdout, so the GPL engine stays out of the MIT
program. The Python package — wheel and sdist — therefore contains only the MIT
part; sidecar/ is excluded from both, and a wheel install falls back to the OS
voice.
A bundle that contains the frozen sidecar is GPL-3.0, not MIT. If you
distribute an installer that ships piper-sidecar beside the app, that
installer is a combined work: ship the GPL-3.0 text and the corresponding source
of the sidecar, and say where to get it. The source repository, the wheel and
the sdist stay MIT. NOTICE records the split.
Two smaller things worth knowing:
- The app icon shown at the top is VistaCreate stock artwork. Its licence allows the picture inside a piece of software but forbids using it as a logo or a trademark, so do not adopt the drawing as your own mark. Replace it with artwork drawn for the app if it ever ships as a named product.
- Never ship a CC BY-NC voice model in a paid build.
- Real-microphone testing for the 27 languages that are wired but untested,
starting with Chinese, Hindi, Gujarati, Georgian, Kyrgyz and Uzbek, whose number
words come from tables in
voice.pyrather thannum2words. - Signed and notarised builds. Today macOS is ad-hoc signed and Windows is unsigned, which is why the table above has a first launch column at all.
- Smaller installers. The Japanese voice needs a 193 MB dictionary
(
sudachidict_core), more than half of the whole download. It could be fetched on first use, the way the voice models already are.