unicode_typing_grader #

English | 한국어 | 日本語

Unicode-aware typing assessment for Dart. Compare what a learner typed with a reference by user-perceived characters, then calculate deterministic edits, accuracy, and gross characters per minute (CPM).

The engine is language-neutral. Its conformance suite includes Korean Hangul, Japanese kana, Latin combining marks, and emoji joined with zero-width joiners.

When would I use this? #

Use this package when your product shows someone a reference and needs to evaluate what they typed. For example:

• a language-learning app that explains exactly which character differed;
• a transcription exercise that measures accuracy and active typing speed;
• a typing assessment that must treat Hangul, kana, combining marks, and emoji as people see them rather than as UTF-16 code units.

It is a grading engine, not a spell checker or autocomplete service. Your app owns the prompt, text field, timer lifecycle, and feedback UI; this package turns the final reference, input, policy, and active duration into reproducible results.

The animation above comes from the browser example, which calls the same gradeTyping API documented below.

Features #

• NFC normalization before comparison
• Extended grapheme cluster comparison instead of UTF-16 code units
• Deterministic insertion, deletion, and substitution paths
• Exact, trim, and collapse whitespace policies
• Exact or ignored terminal punctuation policies
• Integer-backed accuracy and CPM metrics
• Monotonic active-typing timer
• Versioned, language-neutral JSON conformance fixtures
• Pure Dart with no Flutter dependency

Installation #

Add the package from pub.flutter-io.cn:

dependencies:
  unicode_typing_grader: ^0.1.2

Usage #

import 'package:unicode_typing_grader/unicode_typing_grader.dart';

void main() {
  final result = gradeTyping(
    reference: 'Practice makes progress',
    input: 'Practice makes progres',
    policy: TypingGradingPolicy.exact,
    activeElapsed: const Duration(seconds: 30),
  );

  print(result.accuracyPercent); // 95.6
  print(result.charactersPerMinute); // 44.0
  print(result.comparison.deletionCount); // 1
}

The same comparison treats canonically equivalent Latin text as equal:

final comparison = compareNormalizedGraphemes(
  reference: 'é',
  input: 'e\u0301',
);

print(comparison.editDistance); // 0

Comparison policies #

TypingGradingPolicy is explicit and serializable so a client and server can apply the same rules.

All v1 policies normalize text to NFC before applying these rules.

Metric definitions #

Accuracy uses the longer grapheme sequence as its denominator:

accuracy = max(0, max(reference, input) - editDistance)
           / max(reference, input)

accuracyPermille stores this ratio as an integer from 0 to 1000. charactersPerMinuteThousandths stores gross CPM multiplied by 1000. CPM counts input graphemes after policy preparation and does not subtract errors; record raw keystrokes separately if your product needs them. The scaled integers make persisted and cross-language results reproducible.

Policy preparation always runs in this order: NFC normalization, whitespace, then terminal punctuation. A zero active duration produces zero CPM; a negative duration is rejected.

Conformance contract #

conformance/v1/fixtures.json defines observable v1 behavior, including ambiguous edit-path tie breaking. Independent implementations can run the same cases to verify that they agree with the Dart package. See the conformance guide.

The initial contract is tested with characters 1.4.1, based on Unicode 16.0 grapheme segmentation, and unorm_dart 0.3.2, based on Unicode 17.0 normalization data. A dependency update that changes fixture output requires an explicit contract review instead of a silent behavior change.

Performance boundary #

The edit-path algorithm uses O(reference graphemes × input graphemes) time and memory. It is intended for typing prompts and short passages. Applications must limit untrusted input before using it to grade long documents.

Contributing #

See CONTRIBUTING.md. Bug reports and additional Unicode edge cases are welcome, especially examples from scripts not yet represented in the conformance suite.

License #

BSD 3-Clause. See LICENSE.