@devries-cameron20 / t3code Mirror · read-only / commits / 031ee2dd2b
Fork 0

The best way to control Claude Code, Codex and more from any computer or phone

Mirrored from https://github.com/Camo67/t3code.git

Synced Nov 21, 12:15 UTC

fix(server): dictionary spellings apply next to Chinese and Japanese text (#31780)

Co-authored-by: Julius Marminge <julius0216@outlook.com>

Zedong Peng committed Nov 15, 2026 at 11:25 UTC 031ee2dd2b8097f6eb8a8f79cf133356623ef1c7
2 files changed +15 -6
apps/server/src/speech/SpeechService.test.ts
+6 -2
@@ -450,14 +450,18 @@ it.effect("honors exact aliases and phrase boundaries in batch and streamed text
450 { term: "McDonald’s", aliases: [] },
451 { term: "King’s Cross", aliases: [] },
452 { term: "C#", aliases: [] },
453 + { term: "React", aliases: ["瑞安特"] },
454 + { term: "Kubernetes", aliases: [] },
455 ],
456 };
457 const spoken =
458 'a test with a i and ay eye. Robert, Roberts. "the publisher" and the publisher. (git hub), git (hub). ' +
457 - "git hub's API, git hub’s API, git’s hub, mac donald's menu, king's cross. vitamin c and plan c.";
459 + "git hub's API, git hub’s API, git’s hub, mac donald's menu, king's cross. vitamin c and plan c. " +
460 + "我用github和瑞安特,部署到kubernetis上。Robert用git hub。githubで管理する。";
461 const expected =
462 "a test with AI and AI. Bob, Roberts. O'Reilly and the publisher. (GitHub), git (hub). " +
460 - "GitHub's API, GitHub’s API, git’s hub, McDonald’s menu, King’s Cross. vitamin c and plan c.";
463 + "GitHub's API, GitHub’s API, git’s hub, McDonald’s menu, King’s Cross. vitamin c and plan c. " +
464 + "我用GitHub和React,部署到Kubernetes上。Bob用GitHub。GitHubで管理する。";
465 native.transcribe.mockResolvedValueOnce({ text: spoken });
466 native.finish.mockResolvedValueOnce(spoken);
467 expect(yield* speech.transcribe(pcm(), options)).toBe(expected);
apps/server/src/speech/customWords.ts
+9 -4
@@ -1,5 +1,11 @@
1 import type { SpeechCustomWords } from "@t3tools/contracts";
2
3 +// These scripts are written without spaces, so a neighbor in one of them still ends a word.
4 +const UNSPACED_SCRIPTS = "\\p{scx=Han}\\p{scx=Hiragana}\\p{scx=Katakana}";
5 +const LETTER = `[[\\p{L}\\p{M}\\p{N}]--[${UNSPACED_SCRIPTS}]]`;
6 +const WORD_CHARACTER = `[[\\p{L}\\p{M}\\p{N}\\p{Pc}]--[${UNSPACED_SCRIPTS}]]`;
7 +const TOKEN = new RegExp(`${LETTER}(?:[^\\s${UNSPACED_SCRIPTS}]*${LETTER})?`, "gv");
8 +
9 /** Fuzzy-matches preferred terms and rewrites aliases to their preferred spelling. */
10 export function makeSpeechTextCorrector(
11 words: readonly string[],
@@ -112,7 +118,7 @@ export function makeSpeechTextCorrector(
118 const apply = (text: string): string => {
119 if (keys.length === 0) return text;
120 // Leave surrounding punctuation in the source; only horizontal whitespace can join words.
115 - const tokens = [...text.matchAll(/[\p{L}\p{M}\p{N}](?:\S*[\p{L}\p{M}\p{N}])?/gu)]
121 + const tokens = [...text.matchAll(TOKEN)]
122 .filter(([word]) => /^[\p{L}\p{M}\p{N}'’-]+$/u.test(word))
123 .map((match) => ({
124 value: match[0],
@@ -168,10 +174,9 @@ export function makeSpeechTextCorrector(
174 const lookup = new Map(
175 spellings.map(({ spelling, term }) => [spelling.toLocaleLowerCase(), term]),
176 );
171 - const word = "[\\p{L}\\p{M}\\p{N}\\p{Pc}]";
177 const pattern = new RegExp(
173 - `(?<!${word})(?:${spellings.map(({ spelling }) => escapePattern(spelling)).join("|")})(?!${word})`,
174 - "giu",
178 + `(?<!${WORD_CHARACTER})(?:${spellings.map(({ spelling }) => escapePattern(spelling)).join("|")})(?!${WORD_CHARACTER})`,
179 + "giv",
180 );
181 return (text: string) => {
182 let output = "";