diff --git a/CHANGELOG.md b/CHANGELOG.md index 0281cb2..04c3e87 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -13,6 +13,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 - Plugin system for custom pipelines +## [1.16.3] - 2026-11-05 + +> Supersedes 1.16.2, which could rewrite ordinary Korean words in transcripts. + +### Fixed +- Stop the learned-vocabulary pass from rewriting ordinary Korean. Measured against real meeting transcripts it changed 286 of 3,005 words (`하면`→`라면`, `이제`→`이게`, `매일`→`메일`, `정보`→`정도`). Deterministic correction is now limited to your dictionary's pronunciations and the Latin identifiers you type; Korean terms reach recognition through the prompt hint and the polish prompt instead. Re-measured: 0 of 3,005 words changed. +- Drop Hangul-layout typing artifacts (for example `dkssudgktpdy`) from the recognition hint so they no longer bias transcription, and keep the original casing of learned identifiers (`VSCode`). + +### Changed +- Korean terms are no longer auto-replaced word by word. They are used as a recognition hint and as reference for the local model, which decides with sentence context. + ## [1.16.2] - 2026-11-05 > Published from an annotated tag. Installer and update metadata are served by the diff --git a/apps/desktop/tests/main/services/transcript-glossary.test.ts b/apps/desktop/tests/main/services/transcript-glossary.test.ts index f2ce690..8f59ec7 100644 --- a/apps/desktop/tests/main/services/transcript-glossary.test.ts +++ b/apps/desktop/tests/main/services/transcript-glossary.test.ts @@ -24,13 +24,22 @@ function learnPhrase(text: string, times: number): void { } describe('학습한 어휘가 전사 교정에 쓰인다', () => { - it('키보드로 배운 어절과 한 글자 다른 전사 어절을 되돌린다', () => { + it('키보드로 배운 식별자와 한 글자 다른 전사 어절을 되돌린다', () => { + learnPhrase('VSCode 설정 열기', 5) + + const result = getTranscriptGlossaryService().correct('VSCod 설정 열기') + + expect(result.text).toBe('VSCode 설정 열기') + expect(result.corrections).toEqual([{ from: 'VSCod', to: 'VSCode', source: 'learned' }]) + }) + + it('한글 문장은 학습 어휘로 바꾸지 않는다', () => { learnPhrase('회의록 정리해줘', 5) const result = getTranscriptGlossaryService().correct('회이록 정리해줘') - expect(result.text).toBe('회의록 정리해줘') - expect(result.corrections).toEqual([{ from: '회이록', to: '회의록', source: 'learned' }]) + expect(result.text).toBe('회이록 정리해줘') + expect(result.corrections).toEqual([]) }) it('사용자 사전의 발음 표기를 정본 표기로 바꾼다', () => { @@ -53,11 +62,11 @@ describe('학습한 어휘가 전사 교정에 쓰인다', () => { }) it('한 번만 쓴 어절은 학습 어휘로 쓰지 않는다', () => { - learnPhrase('회의록 정리해줘', 1) + learnPhrase('VSCode 설정 열기', 1) - const result = getTranscriptGlossaryService().correct('회이록 정리해줘') + const result = getTranscriptGlossaryService().correct('VSCod 설정 열기') - expect(result.text).toBe('회이록 정리해줘') + expect(result.text).toBe('VSCod 설정 열기') }) it('사전에 단어를 추가하면 캐시를 기다리지 않고 바로 반영된다', () => { diff --git a/docs/map/10-feature-catalog.md b/docs/map/10-feature-catalog.md index ca2267e..520eeb6 100644 --- a/docs/map/10-feature-catalog.md +++ b/docs/map/10-feature-catalog.md @@ -89,7 +89,7 @@ Status quick-reference: `[x]` done+verified · `[~]` partial/unverified · `[ ]` | MTG-11 | Meeting AI chat over transcript | [x] | [ ] | [ ] | [x] | Desktop `MeetingChatPanel` + `MEETING_CHAT` | | MTG-12 | Content reporting for generated docs | [-] | [ ] | [x] | [x] | `content-report` Edge + generation receipts | | MTG-13 | Existing audio/video import into a meeting | [x] | [ ] | [~] | [x] | **2026-11-01:** Desktop `meetingMode:importRecording` reuses `FileTranscriptionService` without duplicate history, saves canonical timestamped transcript + normalized WAV, then opens existing minutes generation. Mobile `MeetingsScreen` creates the workspace then routes `meetingId` + `importRequest` to the existing durable Record meeting pipeline. **2026-11-02:** installed Windows 1.16.0 passed the actual Electron picker→timestamped transcript→local summary/action-item journey, so desktop is `[x]`. Fold6 `.3` import entry/editor rendered, but authenticated device file-to-minutes E2E remains (`GAP-MTG-03`). [Evidence](../evidence/2026-11-02-verification.md). | -| MTG-14 | Learning-informed transcript correction (user dictionary + typed vocabulary) | [~] | [ ] | [ ] | [ ] | **2026-11-05 (root cause + rebuild of the transcript-correction seam):** the learned corpus was only wired to input suggestions. Meeting/live captions ran STT with a hardcoded `다음은 한국어 대화입니다.` prompt (no dictionary, no learned terms), the refine prompt had no glossary, the refine result updated only `CaptionService`'s own copy (meetings and the renderer kept the raw line), and the persisted `rawTranscript` was a plain join of raw segments. Now: core SSOT `@d3ro/core/transcript-correction` (glossary build, dictionary-variant replacement, 1-jamo learned-term correction with a unique-best rule that never touches words the user actually wrote, LLM reference block) + desktop `TranscriptGlossaryService` (dictionary + `personal_phrases`, 15 s cache invalidated on dictionary writes, fail-closed) are wired into `CaptionService` (hint on every recognition, deterministic correction per final line, glossary-aware refine, `segment-updated` event), `MeetingRecordingSession`/`MeetingModeService` (refined lines reach the meeting transcript and renderer via `meetingMode:segmentUpdated`; auto-polish gets the glossary block), `FileTranscriptionService` (per-chunk hint + correction) and `VoiceModeService` (learned-term pass). Tests: core `transcript-correction.test.ts` (16), desktop `transcript-glossary.test.ts` (9); real-app meeting verification still open (`GAP-MTG-04`). | +| MTG-14 | Learning-informed transcript correction (user dictionary + typed vocabulary) | [~] | [ ] | [ ] | [ ] | **2026-11-05 (root cause + rebuild of the transcript-correction seam):** the learned corpus was only wired to input suggestions. Meeting/live captions ran STT with a hardcoded `다음은 한국어 대화입니다.` prompt (no dictionary, no learned terms), the refine prompt had no glossary, the refine result updated only `CaptionService`'s own copy (meetings and the renderer kept the raw line), and the persisted `rawTranscript` was a plain join of raw segments. Now: core SSOT `@d3ro/core/transcript-correction` (glossary build, dictionary-variant replacement, learned-identifier correction with a unique-best rule, LLM reference block) + desktop `TranscriptGlossaryService` (dictionary + `personal_phrases`, 15 s cache invalidated on dictionary writes, fail-closed) are wired into `CaptionService` (hint on every recognition, deterministic correction per final line, glossary-aware refine, `segment-updated` event), `MeetingRecordingSession`/`MeetingModeService` (refined lines reach the meeting transcript and renderer via `meetingMode:segmentUpdated`; auto-polish gets the glossary block), `FileTranscriptionService` (per-chunk hint + correction) and `VoiceModeService` (learned-term pass). **2026-11-05 (v1.16.3 correction policy):** the first cut also rewrote ordinary Hangul words by 1-jamo distance. Measured against the real local corpus (users/_local/d3ro.db, 3 stored meeting transcripts) it changed **286 of 3,005 words** (하면 to 라면, 이제 to 이게, 매일 to 메일, 정보 to 정도). Deterministic correction is now limited to dictionary pronunciations and typed Latin identifiers (canonical casing preserved); Korean terms reach recognition through the STT hint and the polish prompt only. Re-measured on the same corpus: **0 of 3,005**. Real-engine check (sidecar large-v3-turbo, Korean TTS fixture containing the user term): baseline 하람로그에 to hinted transcript haramlog에, and a dictionary pronunciation entry still maps 디에로 to D3RO. Hangul-layout typing artifacts (dkssudgktpdy) are dropped from the hint. Tests: core transcript-correction.test.ts (16), desktop transcript-glossary.test.ts (10); live meeting with real jargon still open (GAP-MTG-04). | --- diff --git a/docs/map/11-gap-backlog.md b/docs/map/11-gap-backlog.md index 430eda8..d86fbf8 100644 --- a/docs/map/11-gap-backlog.md +++ b/docs/map/11-gap-backlog.md @@ -68,7 +68,7 @@ Legend: `[ ]` open · `[~]` in progress · `[!]` blocked externally · `[x]` res | GAP-MTG-01 | Meetings | Audio seek ↔ transcript timestamp sync missing. | `apps/mobile-rn/src/features/meetings/meetings-service.ts`, `MeetingDetailScreen.tsx`, `apps/web/src/components/meetings/{meeting-audio-player,live-transcript-list}.tsx`, `apps/mobile-rn/__tests__/meetings.test.ts` | `[x]` 2026-10-30: 모바일(`MeetingDetailScreen.tsx`)과 웹(`meeting-audio-player.tsx` ↔ `live-transcript-list.tsx`) 및 데스크톱(`TranscriptTab.tsx` / `EditableSegment.tsx`)에 `[MM:SS]` 구간 타임스탬프 파싱·오디오 시크 이동 및 재생 위치별 활성 전사 세그먼트 하이라이트를 구현했다 (`__tests__/meetings.test.ts` 통과). | | GAP-MTG-02 | Meetings | Speaker diarization only partially done on desktop; absent web/mobile. | `docs/phases/phase-15.5-speaker-diarization.md`, SSOT H-014 | Finish desktop pyannote path; expose speaker labels cross-surface. | | GAP-MTG-03 | Meetings | Existing recordings could become plain transcription history, but not a meeting with linked audio and minutes. | desktop `MeetingModeService`, `meeting/imported-recording.ts`, `meeting-mode-handlers.ts`; mobile `MeetingsScreen` → `RecordScreen`; MTG-13 | `[~]` 2026-11-02: installed Windows 1.16.0 passed native picker→timestamped transcript→local summary/action item with a real WAV (`2026-11-02-verification.md`). Desktop GUI gate is complete; authenticated Android picker→upload→transcript→minutes on Fold6 remains. | -| GAP-MTG-04 | Meetings | Learned typing vocabulary never reached meeting transcription: STT used a hardcoded Korean hint, refine had no glossary, refined lines stopped inside `CaptionService`, and the stored transcript was raw segments. | `packages/core/src/transcript-correction.ts`, `apps/desktop/src/main/services/TranscriptGlossaryService.ts`, `CaptionService.ts`, `meeting/MeetingRecordingSession.ts`, `MeetingModeService.ts`, `FileTranscriptionService.ts`, `VoiceModeService.ts`; MTG-14 | `[~]` 2026-11-05: one glossary SSOT wired into caption, meeting, file transcription and dictation; refined lines now propagate to the meeting transcript and renderer (`meetingMode:segmentUpdated`). Evidence: `packages/core/__tests__/transcript-correction.test.ts` 16/16, `apps/desktop/tests/main/services/transcript-glossary.test.ts` 9/9, desktop offline suite, workspace typecheck, touched-file ESLint and `test:hygiene` green. **Remaining:** run a real meeting with the user's actual learned vocabulary and confirm no over-correction of ordinary Korean words, plus a packaged-install check of the new IPC event. | +| GAP-MTG-04 | Meetings | Learned typing vocabulary never reached meeting transcription: STT used a hardcoded Korean hint, refine had no glossary, refined lines stopped inside `CaptionService`, and the stored transcript was raw segments. | `packages/core/src/transcript-correction.ts`, `apps/desktop/src/main/services/TranscriptGlossaryService.ts`, `CaptionService.ts`, `meeting/MeetingRecordingSession.ts`, `MeetingModeService.ts`, `FileTranscriptionService.ts`, `VoiceModeService.ts`; MTG-14 | `[~]` 2026-11-05: one glossary SSOT wired into caption, meeting, file transcription and dictation; refined lines now propagate to the meeting transcript and renderer (`meetingMode:segmentUpdated`). **Direct measurement against the real local database** (`users/_local/d3ro.db`: 1,952 learned phrases / 2,522 typing samples / 3 meeting transcripts) exposed that the first cut also rewrote ordinary Korean (286 of 3,005 words: 하면→라면, 이제→이게, 매일→메일, 정보→정도), and v1.16.3 restricts deterministic correction to dictionary pronunciations plus typed Latin identifiers — re-measured **0 of 3,005**. Real-engine check with the local sidecar (`large-v3-turbo`): a Korean TTS fixture containing the user term transcribes as 하람로그에 with no hint and haramlog에 with the learned hint; a dictionary pronunciation entry maps 디에로 → D3RO. Hangul-layout artifacts (dkssudgktpdy) are excluded from the hint. Evidence: `packages/core/__tests__/transcript-correction.test.ts` 16/16, `apps/desktop/tests/main/services/transcript-glossary.test.ts` 10/10, core 589/589, desktop offline suite, workspace typecheck, touched-file ESLint, `test:hygiene`. **Remaining:** a live meeting with real names/jargon by voice, and a packaged-install check of the new IPC event. | | GAP-INFRA-01 | Build | `apps/mobile-rn` is outside npm workspaces, so root `typecheck`/`test`/`lint` skip it. `typecheck:mobile`/`lint:mobile`/`test:mobile`/`verify:all` root scripts added 2026-09-13 (`package.json`), but membership/CI integration is still open. | `package.json`, `.forgejo/workflows/ci.yml`, `.gitlab-ci.yml` | `[x]` 2026-10-30: React Native Metro/Gradle 네이티브 오토링킹 격리를 위해 `apps/mobile-rn`은 독립 `node_modules`로 유지하고, 루트 `verify:all` 및 `.forgejo/workflows/ci.yml`·`.gitlab-ci.yml`의 `mobile-quality` 잡이 모바일 `lint`·`typecheck`·`test:ci`를 필수 게이트로 실행하도록 정본화했다. **2026-11-04:** root `test:mobile`도 watch-capable 기본 Jest가 아니라 mobile `test:ci`(`jest --runInBand`)를 호출하도록 고정했다. | | GAP-MOB-01 | Legacy | `apps/mobile` Expo skeleton duplicated auth/data code and got version-sync edits. | SSOT G-002 | `[x]` 2026-09-13: deleted `apps/mobile`, removed from `sync-version.mjs` and `package-lock.json`; `version:check` GREEN. | | GAP-INFRA-02 | CI | GitLab admin NAS deploy job is disabled; admin deploy is manual/GHCR. | `.gitlab-ci.yml`, `docs/deployment/nas-deploy-runbook.md` | `[x]` 2026-10-30: 임시 CI DinD 데몬만 바꾸던 구 잡 대신 `docker-compose.nas.yml` 기반의 정본 배포·헬스체크·롤백 절차를 `docs/deployment/nas-deploy-runbook.md`로 문서화하고 `.gitlab-ci.yml` 및 `02-infrastructure.md`에 연결했다. | diff --git a/memory/project_status.md b/memory/project_status.md index ce676f4..a739fb8 100644 --- a/memory/project_status.md +++ b/memory/project_status.md @@ -7,6 +7,7 @@ - 해결: 새 SSOT `packages/core/src/transcript-correction.ts`(글로서리·사전 변형 치환·1자모 학습 어절 교정·LLM 참고 블록) + `TranscriptGlossaryService`(사전+개인 문구 15초 캐시, 사전 변경 시 즉시 무효화, 실패 시 빈 어휘) 를 `CaptionService`(매 인식 힌트·확정 줄 교정·교열 프롬프트·`segment-updated`), `MeetingRecordingSession`/`MeetingModeService`(`meetingMode:segmentUpdated`로 회의 세션·화면 갱신, auto-polish에 용어 블록), `FileTranscriptionService`(조각별 힌트·교정), `VoiceModeService`(학습 어절 패스)에 배선. - 교정 안전 규칙: 사용자가 실제로 쓴 어절은 절대 변경하지 않고, 후보가 동률이면 고치지 않으며, 조사는 그대로 두고 어절만 바꾼다. - 검증: core `transcript-correction.test.ts` 16, 데스크톱 `transcript-glossary.test.ts` 9, 데스크톱 offline 전체(2788)에서 드러난 회귀 2건을 수정해 재통과, workspace typecheck·변경 파일 ESLint·`test:hygiene` GREEN. 실회의 실측은 `GAP-MTG-04`로 남김. +- **1.16.2 배포 직후 실사용 DB로 직접 검증해 결함 발견 → 1.16.3으로 수정:** `users/_local/d3ro.db`(개인 문구 1,952·이핑 표본 2,522·회의 전사 3건)에 결정적 교정을 돌리니 실제 전사 3,005어절 중 **286개가 잘못 바뀌었다**(하면→라면, 이제→이게, 매일→메일, 정보→정도, 일단→일반). 학습 어휘 대부분이 일반 명사·부사·어미라 1자모 거리로는 오인식과 정상 단어를 구분할 수 없다. 수정: 결정적 교정은 **사전 발음 변형 + 사용자가 친 라틴 식별자**만(대소문자 보존), 한글 용어는 STT 힌트·LLM 교열 참고로만. 같은 코퍼스 재측정 **0/3,005**. 실제 사이드카(`large-v3-turbo`)로 한국어 TTS 픽스처 검증: 힌트 없음 `하람로그에` → 힌트 있음 `haramlog에`, 사전 발음 등록 시 `디에로`→`D3RO`. 한영 자판 오타 산물(`dkssudgktpdy`)은 힌트에서 제외. 교훈: **"학습 데이터를 쓴다"는 기능은 실사용 코퍼스로 오탐을 먼저 재고 나서 켠다.** ## v1.9.0 — 전사 구간·프리셋 프롬프트 동기화 (2026-09-27) diff --git a/packages/core/__tests__/transcript-correction.test.ts b/packages/core/__tests__/transcript-correction.test.ts index e6bd0b5..3fed7f3 100644 --- a/packages/core/__tests__/transcript-correction.test.ts +++ b/packages/core/__tests__/transcript-correction.test.ts @@ -90,33 +90,45 @@ describe('replaceDictionaryTerms', () => { describe('correctLearnedTerms', () => { const glossary = buildTranscriptGlossary({ - phrases: phrases(['회의록 정리해줘', 6], ['전사 교정', 4]), + phrases: phrases(['VSCode 설정 열기', 6], ['D3RO Voice 녹음', 4]), }) - it('1자모 차이 나는 전사 어절을 학습 어절로 되돌린다', () => { - const result = correctLearnedTerms('회이록 정리해줘', glossary) + it('학습한 식별자와 한 글자 다른 전사 어절을 되돌린다', () => { + const result = correctLearnedTerms('VSCod 설정 열기', glossary) - expect(result.text).toBe('회의록 정리해줘') - expect(result.corrections).toEqual([{ from: '회이록', to: '회의록', source: 'learned' }]) + expect(result.text).toBe('VSCode 설정 열기') + expect(result.corrections).toEqual([{ from: 'VSCod', to: 'VSCode', source: 'learned' }]) }) - it('조사는 붙인 채로 어절만 고친다', () => { - expect(correctLearnedTerms('회이록은 좋다', glossary).text).toBe('회의록은 좋다') - }) + it('이미 맞게 적힌 식별자는 건드리지 않는다', () => { + const result = correctLearnedTerms('VSCode 설정 열기', glossary) - it('사용자가 실제로 쓴 어절은 건드리지 않는다', () => { - const result = correctLearnedTerms('회의록 정리해줘', glossary) - - expect(result.text).toBe('회의록 정리해줘') + expect(result.text).toBe('VSCode 설정 열기') expect(result.corrections).toEqual([]) }) + it('한글 어절은 1자모만 달라도 고치지 않는다', () => { + // 실사용 코퍼스 측정: 한글까지 고치면 정상 문장이 대량으로 망가졌다(하면→라면, 이제→이게 …). + const korean = buildTranscriptGlossary({ + phrases: phrases(['회의록 정리해줘', 6], ['라면 먹고 싶다', 5]), + }) + + expect(correctLearnedTerms('회이록 정리해줘', korean).text).toBe('회이록 정리해줘') + expect(correctLearnedTerms('하면 좋겠다', korean).text).toBe('하면 좋겠다') + }) + it('후보가 둘 이상이면(동률) 고치지 않는다', () => { const tied = buildTranscriptGlossary({ - phrases: phrases(['정리해', 4], ['정리히', 4]), + phrases: phrases(['vscode 사용', 4], ['vscodf 사용', 4]), }) - expect(correctLearnedTerms('정리하', tied).text).toBe('정리하') + expect(correctLearnedTerms('vscodd', tied).text).toBe('vscodd') + }) + + it('너무 짧은 식별자·한 글자 어절은 대상이 아니다', () => { + const shortTerms = buildTranscriptGlossary({ phrases: phrases(['api 서버', 9]) }) + + expect(correctLearnedTerms('ap 서버', shortTerms).text).toBe('ap 서버') }) it('어휘가 없으면 원문을 그대로 돌려준다', () => { @@ -124,15 +136,6 @@ describe('correctLearnedTerms', () => { expect(result).toEqual({ text: '아무 말이나 적는다', corrections: [] }) }) - - it('사전 정본 표기가 학습 어절보다 우선한다', () => { - const withDictionary = buildTranscriptGlossary({ - dictionary: [{ word: '회의록', pronunciation: null }], - phrases: phrases(['회의녹 정리', 9]), - }) - - expect(correctLearnedTerms('회이록 정리', withDictionary).text).toBe('회의록 정리') - }) }) describe('LLM 교열 참고 블록', () => { diff --git a/packages/core/src/transcript-correction.ts b/packages/core/src/transcript-correction.ts index 981b195..0e06e20 100644 --- a/packages/core/src/transcript-correction.ts +++ b/packages/core/src/transcript-correction.ts @@ -12,7 +12,7 @@ // // 순수 함수만 둔다. DB·Electron 을 모르므로 데스크톱 서비스가 소스를 주입한다. -import { decomposeHangul, editDistance, tokenizeWords } from './typo-correction' +import { editDistance, latinToHangul, tokenizeWords } from './typo-correction' /** 사용자 사전 한 항목 */ export interface GlossaryDictionaryEntry { @@ -59,15 +59,7 @@ const STOP_WORDS: ReadonlySet = new Set([ '이렇게', '그렇게', '저렇게', '여기', '거기', '저기', ]) -/** 토큰에서 떼어 낼 수 있는 조사 (긴 것부터) */ -const JOSA_SUFFIXES: readonly string[] = [ - '이라는', '이라고', '이라도', '에서는', '에게는', '에서도', '으로는', '으로써', - '에서는', '부터', '까지', '에게', '한테', '에서', '으로', '라고', '라는', - '이나', '라도', '이랑', '하고', '이고', '이든', '처럼', '보다', '만큼', - '에는', '에도', '께서', '이란', '이라', - '은', '는', '이', '가', '을', '를', '에', '의', '로', '도', '와', '과', '만', '들', -] - +/** 학습 어휘를 STT 힌트·LLM 교열·식별자 교정에 쓰기 위한 묶음. */ export interface TranscriptGlossary { /** 사전 정본 단어 (원표기) */ terms: readonly string[] @@ -75,6 +67,8 @@ export interface TranscriptGlossary { dictionary: readonly { word: string; variants: readonly string[] }[] /** 소문자 어절 → 빈도 (교정 후보) */ vocabulary: ReadonlyMap + /** 소문자 어절 → 처음 본 표기 (VSCode 같은 대소문자 보존) */ + canonical: ReadonlyMap /** 사용자가 실제로 쓰는 표현 (문체 참고) */ examples: readonly string[] /** STT initial_prompt 힌트 (없으면 빈 문자열) */ @@ -96,6 +90,7 @@ const EMPTY_GLOSSARY: TranscriptGlossary = { terms: [], dictionary: [], vocabulary: new Map(), + canonical: new Map(), examples: [], promptHint: '', } @@ -144,6 +139,7 @@ export function buildTranscriptGlossary(input: { const dictionary: Array<{ word: string; variants: readonly string[] }> = [] const terms: string[] = [] const vocabulary = new Map() + const canonical = new Map() const seen = new Set() for (const entry of (input.dictionary ?? []).slice(0, limits.dictionaryEntries)) { @@ -156,10 +152,13 @@ export function buildTranscriptGlossary(input: { terms.push(word) // 사전 정본은 어떤 학습 어절보다 우선 후보가 되도록 큰 가중치를 준다. bump(vocabulary, key, 1000) + canonical.set(key, word) } const examples: string[] = [] const phraseTokens = new Map() + /** 같은 뜻의 한글 표기가 코퍼스에 있는 영문 토큰(한영 자판 오타 산물) — 힌트·교정에서 뺀다. */ + const hangulObserved = new Set() for (const phrase of (input.phrases ?? []).slice(0, limits.phrases)) { const text = phrase.text?.replace(/\s+/gu, ' ').trim() if (!text) continue @@ -175,7 +174,21 @@ export function buildTranscriptGlossary(input: { for (const token of tokenizeWords(text)) { const word = token.text if (!isLearnableToken(word)) continue - bump(phraseTokens, word.toLowerCase(), weight) + if (/^[가-힣]+$/u.test(word)) hangulObserved.add(word) + const key = word.toLowerCase() + bump(phraseTokens, key, weight) + if (!canonical.has(key)) canonical.set(key, word) + } + } + + // 한영 자판 오타 산물( 같은)은 같은 뜻의 한글 표기가 코퍼스에 있으므로 뺀다. + // 그대로 두면 STT 힌트가 의미 없는 라틴 문자열로 편향되고, 식별자 교정 후보로도 쓰인다. + for (const word of [...phraseTokens.keys()]) { + if (!/^[a-z]+$/u.test(word)) continue + const converted = latinToHangul(word) + if (converted.complete && hangulObserved.has(converted.text)) { + phraseTokens.delete(word) + canonical.delete(word) } } @@ -211,9 +224,9 @@ export function buildTranscriptGlossary(input: { const promptHint = joined ? `용어: ${joined}.` : '' if (vocabulary.size === 0 && terms.length === 0 && examples.length === 0) { - return { ...EMPTY_GLOSSARY, vocabulary: new Map() } + return { ...EMPTY_GLOSSARY, vocabulary: new Map(), canonical: new Map() } } - return { terms, dictionary, vocabulary, examples, promptHint } + return { terms, dictionary, vocabulary, canonical, examples, promptHint } } /** 글로서리에 아무 정보도 없으면 true */ @@ -287,8 +300,8 @@ export function replaceDictionaryTerms( interface VocabularyCandidate { word: string count: number - jamo: readonly string[] - script: 'hangul' | 'latin' + lower: string + chars: readonly string[] } interface VocabularyIndex { @@ -303,11 +316,16 @@ interface VocabularyIndex { */ const vocabularyIndexCache = new WeakMap() -function scriptOf(value: string): 'hangul' | 'latin' | null { - if (/^[가-힣]+$/u.test(value)) return 'hangul' - if (/^[A-Za-z]+$/u.test(value)) return 'latin' - return null -} +/** + * 결정적 교정 대상은 ASCII 식별자(제품명·약어·코드)뿐이다. + * + * 한글 어절을 「학습 어휘와 1자모 다르면 교체」로 고치는 방식은 실사용 DB 코퍼스로 측정했을 때 + * 정상 문장을 대량으로 망쳤다(회의 전사 3,005어절 중 286개 변경: 하면→라면, 이제→이게, 매일→메일, + * 정보→정도, 일단→일반 …). 사용자가 실제로 쓰는 어휘에는 일반 명사·부사·어미가 대부분이라 + * 1자모 거리만으로는 「오인식」과 「정상 단어」를 구분할 수 없다. 한글 용어는 STT 힌트와 LLM 교열에만 + * 쓰고, 결정적으로는 고치지 않는다(정밀도 우선). + */ +const IDENTIFIER_TOKEN = /^[A-Za-z][A-Za-z0-9]{3,}$/u function buildVocabularyIndex(glossary: TranscriptGlossary): VocabularyIndex { const cached = vocabularyIndexCache.get(glossary) @@ -321,40 +339,37 @@ function buildVocabularyIndexUncached(glossary: TranscriptGlossary): VocabularyI const byLength = new Map() const known = new Set() for (const [word, count] of glossary.vocabulary) { - const script = scriptOf(word) - if (!script) continue + if (!IDENTIFIER_TOKEN.test(word)) continue known.add(word) const candidate: VocabularyCandidate = { - word, + word: glossary.canonical.get(word) ?? word, count, - jamo: decomposeHangul(word), - script, + lower: word.toLowerCase(), + chars: Array.from(word.toLowerCase()), } - const bucket = byLength.get(candidate.jamo.length) + const bucket = byLength.get(candidate.chars.length) if (bucket) bucket.push(candidate) - else byLength.set(candidate.jamo.length, [candidate]) + else byLength.set(candidate.chars.length, [candidate]) } return { byLength, known } } /** - * 학습 어절에서 1자모 차이 나는 유일한 후보를 찾는다. + * 학습 어휘에서 1글자 차이 나는 유일한 후보를 찾는다. * 동률이면(어느 쪽이 맞는지 알 수 없으면) 고치지 않는다. */ function bestCandidate(token: string, index: VocabularyIndex): VocabularyCandidate | null { - const script = scriptOf(token) - if (!script) return null - const jamo = decomposeHangul(token) + const lower = token.toLowerCase() + const chars = Array.from(lower) let best: VocabularyCandidate | null = null let bestCount = 0 let ties = 0 - for (const length of [jamo.length - 1, jamo.length, jamo.length + 1]) { + for (const length of [chars.length - 1, chars.length, chars.length + 1]) { const bucket = index.byLength.get(length) if (!bucket) continue for (const candidate of bucket) { - if (candidate.script !== script) continue - if (candidate.jamo.length === jamo.length && candidate.word === token.toLowerCase()) continue - if (editDistance(jamo, candidate.jamo) > 1) continue + if (candidate.lower === lower) continue + if (editDistance(chars, candidate.chars) > 1) continue if (candidate.count > bestCount) { best = candidate bestCount = candidate.count @@ -367,27 +382,9 @@ function bestCandidate(token: string, index: VocabularyIndex): VocabularyCandida return ties === 0 ? best : null } -interface TokenAttempt { - /** 비교할 어절 (조사를 떼어 낸 형태일 수 있다) */ - text: string -} - -function tokenAttempts(token: string): TokenAttempt[] { - const attempts: TokenAttempt[] = [{ text: token }] - for (const josa of JOSA_SUFFIXES) { - if (token.length - josa.length < 2) continue - if (!token.endsWith(josa)) continue - const stem = token.slice(0, token.length - josa.length) - if (!/^[가-힣A-Za-z0-9]+$/u.test(stem)) continue - attempts.push({ text: stem }) - break - } - return attempts -} - /** - * 학습 어휘로 어절을 되돌린다. 띄어쓰기·문장부호·조사는 그대로 두고 어절만 바꾼다. - * 사용자가 실제로 쓴 어절(어휘에 있음)은 절대 건드리지 않는다. + * 학습 어휘로 ASCII 식별자 어절을 되돌린다 (예: `vscod` → `VSCode`). + * 한글 어절·띄어쓰기·문장부호는 건드리지 않는다. */ export function correctLearnedTerms( text: string, @@ -397,26 +394,20 @@ export function correctLearnedTerms( if (!trimmed || glossary.vocabulary.size === 0) return { text: trimmed, corrections: [] } const index = buildVocabularyIndex(glossary) + if (index.byLength.size === 0) return { text: trimmed, corrections: [] } + const corrections: TranscriptCorrectionEntry[] = [] let out = '' let cursor = 0 for (const token of tokenizeWords(trimmed)) { - let replaced: { value: string; prefixLength: number } | null = null - for (const attempt of tokenAttempts(token.text)) { - if (index.known.has(attempt.text.toLowerCase())) continue - if (attempt.text.length < 2) continue - const candidate = bestCandidate(attempt.text, index) - if (!candidate) continue - replaced = { value: candidate.word, prefixLength: attempt.text.length } - break - } - if (!replaced) continue - const original = token.text.slice(0, replaced.prefixLength) - if (original === replaced.value) continue - corrections.push({ from: original, to: replaced.value, source: 'learned' }) - out += trimmed.slice(cursor, token.start) + replaced.value - cursor = token.start + replaced.prefixLength + if (!IDENTIFIER_TOKEN.test(token.text)) continue + if (index.known.has(token.text.toLowerCase())) continue + const candidate = bestCandidate(token.text, index) + if (!candidate) continue + corrections.push({ from: token.text, to: candidate.word, source: 'learned' }) + out += trimmed.slice(cursor, token.start) + candidate.word + cursor = token.end } if (corrections.length === 0) return { text: trimmed, corrections: [] }