104 ·GitHub
What is this?
A zero-dependency JavaScript library that matches a Chinese name heard through speech recognition (or any garbled transcript) to the correct entry in a roster, comparing by pinyin syllables and surfacing candidates with reasons instead of forcing a decision when it is unsure.
A zero-dependency JavaScript library that matches a Chinese name heard through speech recognition (or any garbled transcript) to the correct entry in a roster, comparing by pinyin syllables and surfacing candidates with reasons instead of forcing a decision when it is unsure.Is this a full speech-to-text or attendance solution?No, it is only the name-matching component for developers. It does not do speech recognition, does not extract scores, and does not write anything to a database — those parts are left to the integrator.
No, it is only the name-matching component for developers. It does not do speech recognition, does not extract scores, and does not write anything to a database — those parts are left to the integrator.How do I use it in a project?Just require() it directly in Node.js 14+ since it has zero runtime dependencies and ships as CommonJS; to use it in a browser you first need to bundle it with esbuild, webpack, or vite.
Just require() it directly in Node.js 14+ since it has zero runtime dependencies and ships as CommonJS; to use it in a browser you first need to bundle it with esbuild, webpack, or vite.Do I need Python, an API key, or any external service to run it?No. It is pure rule-based matching with no dependencies, no network calls, and no API keys; everything runs locally and instantly.
No. It is pure rule-based matching with no dependencies, no network calls, and no API keys; everything runs locally and instantly.How reliable is a match?Matches fall into eight tiers (exact, samePinyin, oneLetter, surnameNear, accent, sameChars, loneSurname, nearPinyin); only the top four tiers can be marked 'confident', and a high numeric score on a lower tier is still just a guess, not a verified identity.
Matches fall into eight tiers (exact, samePinyin, oneLetter, surnameNear, accent, sameChars, loneSurname, nearPinyin); only the top four tiers can be marked 'confident', and a high numeric score on a lower tier is still just a guess, not a verified identity.Can it also split one long spoken sentence into separate per-person records?Yes, a helper splits a transcript whenever a digit is immediately followed by a Chinese character, treating that as a new person, while a following '+', '.', or another digit means the score still belongs to the same person; it deliberately leaves Chinese-written numbers or addition/subtraction phrasing unsplit rather than guessing.
Yes, a helper splits a transcript whenever a digit is immediately followed by a Chinese character, treating that as a new person, while a following '+', '.', or another digit means the score still belongs to the same person; it deliberately leaves Chinese-written numbers or addition/subtraction phrasing unsplit rather than guessing.What should I check before deploying it on my real roster?Run the matcher once against your actual name list locally to find any characters missing from the built-in pinyin table and add them to pinyin-table.js — a missing character fails to match silently, without throwing an error.
Run the matcher once against your actual name list locally to find any characters missing from the built-in pinyin table and add them to pinyin-table.js — a missing character fails to match silently, without throwing an error.Why not just use a generic fuzzy string-matching library for Chinese names?Generic libraries compare character similarity, but Chinese names need to be compared by syllable: homophone characters read identically, some surnames change pronunciation (单/曾/区/仇/解/查), and concatenated pinyin strings can collide between unrelated names, all of which a plain edit-distance approach gets wrong.
Generic libraries compare character similarity, but Chinese names need to be compared by syllable: homophone characters read identically, some surnames change pronunciation (单/曾/区/仇/解/查), and concatenated pinyin strings can collide between unrelated names, all of which a plain edit-distance approach gets wrong.Should I bolt an LLM on top for the tricky cases?Only for things rules cannot catch, like self-corrections mid-sentence; the author recommends a non-reasoning model since reasoning models burn the token budget on hidden thinking and can return empty responses, keeping all score math local and still passing the model's output through this library's own confidence check before trusting it.
Only for things rules cannot catch, like self-corrections mid-sentence; the author recommends a non-reasoning model since reasoning models burn the token budget on hidden thinking and can return empty responses, keeping all score math local and still passing the model's output through this library's own confidence check before trusting it.What are similar projects?lieflat-less-ai-tone, Fuse, flexsearch, Llama-Chinese, Chinese-Word-Vectors, chinese-poetry, how-to-live-better, natural
lieflat-less-ai-tone, Fuse, flexsearch, Llama-Chinese, Chinese-Word-Vectors, chinese-poetry, how-to-live-better, naturalWhich topics is it in?chinese-nlp, cjk, fuzzy-search, data-processing, text-processing, zero-dependency, algorithm, chinese-language, natural-language-processing, offline-first, nlp, javascript-library
chinese-nlp, cjk, fuzzy-search, data-processing, text-processing, zero-dependency, algorithm, chinese-language, natural-language-processing, offline-first, nlp, javascript-library