Plan In App Readings

Plan: readings made in the app, for sentences the learner adds

Source of truth: this wiki page.

Written 2026-10-05.

What the owner asked

The app lacks built-in script-to-Latin transliteration since that's handled by the Python tool at deck creation time: can this be rule generated?? with how much accuracy?

if accurate enough, implement

Then, before anything was built:

this is a plan, not something for you to do. all decks will still get transliteration from wikitionary or soemthing, only user added sentences will get in app transliteration for training on those sentences

Asked earlier which languages, and how readings below the bar are shown, the owner chose:

• All seven Indian languages.
• Readings in a language below 95% are shown as a suggestion to check or edit.

Scope

• Course decks get their readings from Wiktionary or a similar source (tools/import_wiktionary.py, #518), not from the app.
• The app makes readings only for sentences the learner adds:
  • text imported (import-text.md);
  • cards of the learner's own, for training on them.

How accurate rules are: measured

tools/transcribe.py already makes readings by rule. Normally it takes a hint, the word as people type it, to tell which vowels are said. The app has no hint for a learner's sentence. So it was run with no hint on every card and grammar form of the seven Indian languages, 7,010 readings, and compared with the decks' readings.

Metric: a reading counts as right when it equals the deck's exactly. The table also shows:

• the share right once accents are ignored, as a typed answer is compared;
• the character error rate.

Language	Right as it stands	With the three fixes below	Ignoring accents	Character errors
Telugu	98.5%	98.5%	98.5%	0.37%
Kannada	98.2%	98.2%	98.2%	0.44%
Hindi	97.2%	97.9%	97.9%	0.29%
Gujarati	94.1%	95.6%	95.6%	0.67%
Marathi	88.8%	90.7%	90.7%	1.61%
Assamese	82.0%	87.2%	87.2%	1.89%
Bengali	67.5%	86.9%	88.2%	2.25%
All	87.9%	93.0%	93.2%	1.12%

The three fixes apply only when there is no hint:

1. A word's last vowel after a cluster ending in r, y, v, n, ṇ, m or l is said (Hindi, Marathi, Gujarati): मित्र mitra, पत्र patra.
2. Bengali drops an inherent vowel inside a word as Hindi does, between a vowel and a consonant that has a vowel after it: আপনার āpnār.
3. Sibilants in Bengali and Assamese follow the language's table, so Bengali ś and Assamese x, except before a dental or a stop, where they are s. With no hint, the tool used its own typed spelling s as if it were a hint, and wrote s for every স.

Caveats (confidence medium):

• The deck readings were made by the same tool with a hint, and no speaker has checked them. These figures measure agreement with the course, not with speech.
• The fixes were found on the same words they were measured on, so new sentences will score lower.

What is still wrong:

• vowels heard both ways in Hindi and Marathi (janvarī or janvrī);
• English loans spelled with फ but said with f;
• Bengali's inherent o, which needs a word list as much as rules.

What it takes

1. The three fixes in tools/transcribe.py, in its no-hint path only, so the tool and the app agree. Readings made with a hint do not change.
2. A Dart port of the no-hint path, in lib/core, free of Flutter:
   • the script tables, the units, the clusters;
   • the rules above, and the tool's short list of words said otherwise;
   • readings only. The IPA can follow the same way.
3. Tests:
   • the Dart output equals the Python tool's on a fixture the tool writes, about 100 readings per language;
   • each language's agreement with the decks stays at or above the figures above.
4. Shown as made: a reading the app made is marked as such. In Marathi, Assamese and Bengali, below 95%, it is a suggestion the learner checks or edits before it is used, as the owner chose.

To decide

• Whether the IPA is made in the app too.
• Whether a reading the learner corrects is kept as theirs, and offered as a fix to the course.

Estimate

• The fixes in the Python tool, with tests: about 2 hours.
• The Dart port, with the parity fixture and accuracy tests: about 6–8 hours.
• Marking and editing in the screens that use it: 1–2 hours.

Confidence: medium.

See also: Roadmap · Plans-Index · Home