feat(analysis): support voice transcriptions in ChatLab

适配导入文件中的语音转写内容，让 ChatLab 在语言分析、词频统计和口头禅分析中使用实际转写文字，而不是把语音时长标签当作文本。

识别并移除语音消息前的时长标签，保留转写内容；纯语音占位符继续从文本分析中排除。

统一桌面端、CLI Web 和 Web WASM 使用的文本清洗逻辑，并补充 ChatLab JSONL 导入、词频、语言偏好和口头禅分析测试。

测试使用脱敏后的合成消息，不包含个人聊天记录。

packages/core/src/nlp/index.ts

@@ -24,4 +24,10 @@ export { POS_TAG_DEFINITIONS, MEANINGFUL_POS_TAGS } from './pos-tags'
 
 export { CHINESE_STOPWORDS, ENGLISH_STOPWORDS, JAPANESE_STOPWORDS, getStopwords, isStopword } from './stopwords'
 
-export { cleanText, isValidWord } from './text-utils'
+export {
+  cleanText,
+  isMediaPlaceholderContent,
+  isValidWord,
+  stripMediaPlaceholders,
+  stripVoiceTranscriptionPrefix,
+} from './text-utils'

packages/core/src/nlp/text-utils.test.ts

@@ -1,6 +1,6 @@
 import { describe, it } from 'node:test'
 import assert from 'node:assert/strict'
-import { cleanText } from './text-utils'
+import { cleanText, isMediaPlaceholderContent, stripVoiceTranscriptionPrefix } from './text-utils'
 
 describe('cleanText', () => {
   const cases = [
@@ -22,4 +22,14 @@ describe('cleanText', () => {
       assert.equal(cleanText(input), expected)
     })
   }
+
+  it('removes duration-bearing voice labels while retaining the transcription', () => {
+    assert.equal(cleanText('[语音 3秒] 这是测试语音内容。'), '这是测试语音内容')
+    assert.equal(stripVoiceTranscriptionPrefix('[语音 3秒] 这是测试语音内容。'), '这是测试语音内容。')
+  })
+
+  it('recognizes duration-bearing media-only content as a placeholder', () => {
+    assert.equal(isMediaPlaceholderContent('[语音 3秒]'), true)
+    assert.equal(isMediaPlaceholderContent('[语音 3秒] 有转写内容'), false)
+  })
 })

packages/core/src/nlp/text-utils.ts

@@ -8,8 +8,21 @@ const PUNCTUATION_REGEX = /[!"#$%&'()*+,\-./:;<=>?@[\\\]^_`{|}~，。！？、
 const URL_REGEX = /https?:\/\/[^\s]+/g
 const MENTION_REGEX = /@[^\s@]+/g
 const PURE_NUMBER_REGEX = /^\d+$/
-const SYSTEM_PLACEHOLDER_REGEX =
-  /\[(?:图片|视频|语音|文件|动画表情|表情|链接|位置|名片|红包|转账|音乐|Image|Video|Voice|File|Sticker|Link)\]/gi
+const MEDIA_PLACEHOLDER_NAME_PATTERN =
+  '(?:图片|视频|语音|文件|动画表情|表情|链接|位置|地理位置|名片|红包|转账|音乐|Image|Photo|Video|Voice|Audio|File|Sticker|Link|Location)'
+const MEDIA_PLACEHOLDER_DURATION_PATTERN = '(?:\\s+\\d+(?:\\.\\d+)?\\s*(?:秒|s))?'
+const SYSTEM_PLACEHOLDER_REGEX = new RegExp(
+  `\\[${MEDIA_PLACEHOLDER_NAME_PATTERN}${MEDIA_PLACEHOLDER_DURATION_PATTERN}\\]`,
+  'gi'
+)
+const MEDIA_PLACEHOLDER_ONLY_REGEX = new RegExp(
+  `^\\[${MEDIA_PLACEHOLDER_NAME_PATTERN}${MEDIA_PLACEHOLDER_DURATION_PATTERN}\\]$`,
+  'i'
+)
+const VOICE_TRANSCRIPTION_PREFIX_REGEX = new RegExp(
+  `^\\s*\\[(?:语音|Voice|Audio)${MEDIA_PLACEHOLDER_DURATION_PATTERN}\\]\\s*`,
+  'i'
+)
 const BRACKET_EMOJI_PLACEHOLDER_REGEX =
   /(?:\[|【)([\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}\p{Letter}\p{Number}_-]{1,16})(?:\]|】)/gu
 const CJK_TEXT_REGEX = /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}]/u
@@ -118,6 +131,21 @@ export function cleanText(text: string): string {
     .trim()
 }
 
+/** Remove media placeholders, including duration-bearing labels such as `[语音 3秒]`. */
+export function stripMediaPlaceholders(text: string): string {
+  return text.replace(SYSTEM_PLACEHOLDER_REGEX, ' ')
+}
+
+/** Return the transcribed text after a leading voice-message label. */
+export function stripVoiceTranscriptionPrefix(text: string): string {
+  return text.replace(VOICE_TRANSCRIPTION_PREFIX_REGEX, '').trim()
+}
+
+/** Whether the complete value is a media placeholder rather than message text. */
+export function isMediaPlaceholderContent(text: string): boolean {
+  return MEDIA_PLACEHOLDER_ONLY_REGEX.test(text.trim())
+}
+
 /**
  * 判断是否为有效词语
  */

packages/core/src/query/__tests__/browser-word-frequency.test.ts

@@ -119,6 +119,25 @@ describe('getBrowserWordFrequency', () => {
     assert.equal(result.uniqueWords, 2)
   })
 
+  it('keeps voice transcription words but removes the duration label', () => {
+    const insert = raw.prepare('INSERT INTO message (id, sender_id, ts, type, content) VALUES (?, ?, ?, ?, ?)')
+    insert.run(6, 1, 600, 0, '[Voice 3s] sample transcript')
+    insert.run(7, 1, 700, 0, '[Voice 3s] sample transcript')
+
+    const result = getBrowserWordFrequency(database, {
+      sessionId: 'session-one',
+      locale: 'en-US',
+      topN: 20,
+      minCount: 2,
+      posFilterMode: 'all',
+      enableStopwords: false,
+    })
+
+    const words = result.words.map((word) => word.word)
+    assert.ok(words.includes('sample'))
+    assert.ok(!words.includes('voice'))
+  })
+
   it('applies member, time, excluded-word, and excluded-message filters', () => {
     const result = getBrowserWordFrequency(database, {
       sessionId: 'session-one',

packages/core/src/query/advanced/languagePreference.ts

@@ -7,6 +7,7 @@
 
 import type { TimeFilter } from '@openchatlab/shared-types'
 import type { DatabaseAdapter } from '../../interfaces'
+import { stripMediaPlaceholders, stripVoiceTranscriptionPrefix } from '../../nlp'
 import { buildTimeFilter } from '../filters'
 import { isSystemPlaceholderContent } from './text-filters'
 
@@ -46,7 +47,7 @@ const RE_PUNCTUATION = /[!"#$%&'()*+,\-./:;<=>?@[\\\]^_`{|}~，。！？、；
 const RE_PURE_NUMBER = /^\d+$/
 
 function cleanTextForNlp(text: string): string {
-  return text
+  return stripMediaPlaceholders(text)
     .replace(RE_URL, ' ')
     .replace(RE_MENTION, ' ')
     .replace(RE_EMOJI, ' ')
@@ -106,14 +107,15 @@ export function getLanguagePreferenceAnalysis(db: DatabaseAdapter, params: Langu
 
   const memberMessages = new Map<number, { name: string; messages: string[] }>()
   for (const row of rows) {
-    if (isSystemPlaceholderContent(row.content)) continue
+    const content = stripVoiceTranscriptionPrefix(row.content)
+    if (!content || isSystemPlaceholderContent(content)) continue
 
     let entry = memberMessages.get(row.memberId)
     if (!entry) {
       entry = { name: row.name, messages: [] }
       memberMessages.set(row.memberId, entry)
     }
-    entry.messages.push(row.content)
+    entry.messages.push(content)
   }
 
   const isChinese = locale.startsWith('zh')

packages/core/src/query/advanced/repeat.test.ts

@@ -170,6 +170,30 @@ describe('getCatchphraseAnalysis', () => {
     )
     db.close()
   })
+
+  it('shows voice transcriptions without the duration label', () => {
+    const db = createCatchphraseDb([
+      { memberId: 1, content: '[语音 3秒] 测试转写内容', count: 3 },
+      { memberId: 1, content: '[语音 2秒]', count: 5 },
+    ])
+
+    const result = getCatchphraseAnalysis(db)
+
+    assert.deepEqual(result.members[0]?.catchphrases, [{ content: '测试转写内容', count: 3 }])
+    db.close()
+  })
+
+  it('aggregates the same transcription when voice durations differ', () => {
+    const db = createCatchphraseDb([
+      { memberId: 1, content: '[语音 2秒] 同一句测试转写', count: 1 },
+      { memberId: 1, content: '[语音 3秒] 同一句测试转写', count: 1 },
+    ])
+
+    const result = getCatchphraseAnalysis(db)
+
+    assert.deepEqual(result.members[0]?.catchphrases, [{ content: '同一句测试转写', count: 2 }])
+    db.close()
+  })
 })
 
 describe('getLanguagePreferenceAnalysis', () => {
@@ -188,4 +212,28 @@ describe('getLanguagePreferenceAnalysis', () => {
     assert.equal(result.members[0]?.totalMessages, 2)
     assert.deepEqual(result.members[0]?.catchphrases, [{ content: 'noted', count: 2 }])
   })
+
+  it('uses voice transcription text for language analysis instead of the duration label', () => {
+    const db = createRowsDb([
+      { memberId: 1, name: 'Alice', content: '[Voice 3s] sample transcript' },
+      { memberId: 1, name: 'Alice', content: '[Voice 3s] sample transcript' },
+      { memberId: 1, name: 'Alice', content: '[语音 2秒]' },
+      { memberId: 2, name: 'Bob', content: '普通文本' },
+      { memberId: 2, name: 'Bob', content: '普通文本' },
+    ])
+
+    const result = getLanguagePreferenceAnalysis(db, { locale: 'en-US' })
+    const alice = result.members.find((member: { name: string }) => member.name === 'Alice')
+
+    assert.equal(alice?.totalMessages, 2)
+    assert.deepEqual(alice?.catchphrases, [{ content: 'sample transcript', count: 2 }])
+    assert.equal(
+      alice?.topWords.some((word: { word: string }) => word.word === 'voice'),
+      false
+    )
+    assert.equal(
+      alice?.topWords.some((word: { word: string }) => word.word === 'sample'),
+      true
+    )
+  })
 })

packages/core/src/query/advanced/repeat.ts

@@ -4,9 +4,21 @@
 
 import type { TimeFilter } from '@openchatlab/shared-types'
 import type { DatabaseAdapter } from '../../interfaces'
+import { stripVoiceTranscriptionPrefix } from '../../nlp'
 import { buildTimeFilter } from '../filters'
 import { isSystemPlaceholderContent } from './text-filters'
 
+// Keep one-off voice transcription rows in the grouped query so JavaScript can
+// normalize away duration labels before applying the repeated-phrase threshold.
+const VOICE_TRANSCRIPTION_SQL = `
+        TRIM(msg.content) LIKE '[语音] %' OR
+        TRIM(msg.content) LIKE '[语音 %] %' OR
+        TRIM(msg.content) LIKE '[Voice] %' OR
+        TRIM(msg.content) LIKE '[Voice %] %' OR
+        TRIM(msg.content) LIKE '[Audio] %' OR
+        TRIM(msg.content) LIKE '[Audio %] %'
+      `
+
 export interface CatchphraseItem {
   content: string
   count: number
@@ -48,7 +60,7 @@ export function getCatchphraseAnalysis(db: DatabaseAdapter, filter?: TimeFilter)
       JOIN member m ON msg.sender_id = m.id
       ${whereClause}
       GROUP BY m.id, TRIM(msg.content)
-      HAVING COUNT(*) >= 2
+      HAVING COUNT(*) >= 2 OR (${VOICE_TRANSCRIPTION_SQL})
       ORDER BY m.id, count DESC
       `
     )
@@ -61,9 +73,11 @@ export function getCatchphraseAnalysis(db: DatabaseAdapter, filter?: TimeFilter)
   }>
 
   const memberMap = new Map<number, MemberCatchphrase>()
+  const phraseMaps = new Map<number, Map<string, CatchphraseItem>>()
 
   for (const row of rows) {
-    if (isSystemPlaceholderContent(row.content)) continue
+    const content = stripVoiceTranscriptionPrefix(row.content)
+    if (!content || isSystemPlaceholderContent(content)) continue
 
     if (!memberMap.has(row.memberId)) {
       memberMap.set(row.memberId, {
@@ -74,13 +88,23 @@ export function getCatchphraseAnalysis(db: DatabaseAdapter, filter?: TimeFilter)
       })
     }
 
-    const member = memberMap.get(row.memberId)!
-    if (member.catchphrases.length < 100) {
-      member.catchphrases.push({ content: row.content, count: row.count })
+    let phrases = phraseMaps.get(row.memberId)
+    if (!phrases) {
+      phrases = new Map<string, CatchphraseItem>()
+      phraseMaps.set(row.memberId, phrases)
     }
+    const existing = phrases.get(content)
+    if (existing) existing.count += Number(row.count)
+    else phrases.set(content, { content, count: Number(row.count) })
   }
 
   const members = Array.from(memberMap.values())
+  for (const member of members) {
+    member.catchphrases = [...(phraseMaps.get(member.memberId)?.values() ?? [])]
+      .filter((item) => item.count >= 2)
+      .sort((a, b) => b.count - a.count || a.content.localeCompare(b.content))
+      .slice(0, 100)
+  }
   members.sort((a, b) => {
     const countDifference = (b.catchphrases[0]?.count ?? 0) - (a.catchphrases[0]?.count ?? 0)
     return countDifference || a.memberId - b.memberId

packages/core/src/query/advanced/text-filters.ts

@@ -1,3 +1,5 @@
+import { isMediaPlaceholderContent } from '../../nlp'
+
 const SYSTEM_PLACEHOLDER_CONTENTS = new Set([
   '[图片]',
   '[视频]',
@@ -24,5 +26,6 @@ const SYSTEM_PLACEHOLDER_CONTENTS = new Set([
 ])
 
 export function isSystemPlaceholderContent(content: string): boolean {
-  return SYSTEM_PLACEHOLDER_CONTENTS.has(content.trim())
+  const trimmed = content.trim()
+  return SYSTEM_PLACEHOLDER_CONTENTS.has(trimmed) || isMediaPlaceholderContent(trimmed)
 }

packages/parser/src/formats/chatlab-jsonl.test.ts

@@ -84,6 +84,45 @@ test('ChatLab JSONL emits message batches and progress while reading', async ()
   }
 })
 
+test('preserves a duration-bearing voice transcription in the parsed message content', async () => {
+  const root = makeTempDir()
+  const filePath = path.join(root, 'voice-transcription.jsonl')
+  const transcription = '[语音 3秒] 这是测试语音内容。'
+  fs.writeFileSync(
+    filePath,
+    [
+      JSON.stringify({
+        _type: 'header',
+        chatlab: { version: '0.0.2', exportedAt: 1711468800 },
+        meta: { name: 'Voice transcription', platform: 'douyin', type: 'private' },
+      }),
+      JSON.stringify({ _type: 'member', platformId: 'member-1', accountName: 'Alice' }),
+      JSON.stringify({
+        _type: 'message',
+        sender: 'member-1',
+        accountName: 'Alice',
+        timestamp: 1711468800,
+        type: MessageType.TEXT,
+        content: transcription,
+        platformMessageId: 'voice-1',
+      }),
+    ].join('\n') + '\n',
+    'utf-8'
+  )
+
+  try {
+    const messages: Array<{ type: number; content: string | null }> = []
+    for await (const event of parseFileWithFormat('chatlab-jsonl', { filePath })) {
+      if (event.type === 'messages') {
+        messages.push(...event.data.map(({ type, content }) => ({ type, content })))
+      }
+    }
+    assert.deepEqual(messages, [{ type: MessageType.TEXT, content: transcription }])
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
 test('directory entry detection accepts a top-level ChatLab JSONL file', () => {
   const root = makeTempDir()
   const filePath = path.join(root, 'chat.jsonl')