feat: Phase 15.5-2 — 오디오 보존 + pyannote 실제 화자 구분

녹음 중 PCM 데이터를 WAV 파일로 저장 (meeting-audio/ 디렉토리).
diarizeSession에서 WAV 존재 + HF 토큰 있으면 sidecar /diarize 호출.
pyannote 실패 시 LLM 폴백. 전사 세그먼트-화자 시간 매칭.

CLAUDE.md

@@ -32,7 +32,7 @@ npm run typecheck  # tsc --noEmit
 - 녹음 UI: 9개 웨이브 바, cos 분포, 100ms
 
 ## 현재 상태
-Phase 1~15 전체 완료 + 랜딩 페이지(site/).
+Phase 1~15.5 전체 완료 + 랜딩 페이지(site/).
 차단 이슈: @nut-tree-fork/nut-js 포크 사용
 
 ## 설계 문서 (구현 시 Read 도구로 참조)
@@ -52,6 +52,7 @@ Phase13+: VoiceConversationService, RAGService, VoiceActionService
 Phase14: MeetingModeService
 Phase14.5: MeetingDocTemplateService + 다중 문서 생성/편집/내보내기
 Phase15: Auto Polish + AI 채팅 + 마인드맵 + 공유
+Phase15.5: 화자 구분 (LLM 추정 + pyannote 준비)
 
 ## DS 컴포넌트 (src/renderer/components/ds/)
 CrtDisplay, InstrumentPanel, Led, PhysicalButton, MetalCard,

src/main/services/MeetingModeService.ts

@@ -65,6 +65,10 @@ class MeetingModeService extends EventEmitter {
   /** CaptionService session-saved 방지 플래그 */
   private _meetingModeActive = false
 
+  /** Phase 15.5-2: 녹음 중 오디오 WAV 파일 보존 */
+  private _audioBuffersForFile: Buffer[] = []
+  private _audioFilePath: string | null = null
+
   // Phase 15: Meeting AI Chat
   private _chatHistory: Array<{ role: 'user' | 'assistant'; content: string }> = []
   private _chatAbortController: AbortController | null = null
@@ -179,8 +183,11 @@ class MeetingModeService extends EventEmitter {
     // 오디오 레벨 모니터링 시작
     const { getAudioCaptureService, calculateRMS } = await import('./AudioCaptureService')
     const audioCaptureService = getAudioCaptureService()
+    this._audioBuffersForFile = []
     this._audioDataHandler = (payload: { buffer: Buffer; timestamp: number }) => {
       this._lastRms = calculateRMS(payload.buffer)
+      // Phase 15.5-2: 오디오 버퍼 축적 (WAV 저장용)
+      this._audioBuffersForFile.push(Buffer.from(payload.buffer))
     }
     audioCaptureService.on('audio-data', this._audioDataHandler)
 
@@ -305,6 +312,23 @@ class MeetingModeService extends EventEmitter {
         updatedAt: now,
       }).where(eq(meetingSessions.id, sessionId)).run()
 
+      // Step 2.5: 오디오 WAV 파일 저장 (화자 구분용)
+      if (this._audioBuffersForFile.length > 0) {
+        try {
+          const { app } = await import('electron')
+          const path = await import('path')
+          const audioDir = path.join(app.getPath('userData'), 'meeting-audio')
+          if (!fs.existsSync(audioDir)) fs.mkdirSync(audioDir, { recursive: true })
+          const wavPath = path.join(audioDir, `${sessionId}.wav`)
+          this._saveWav(wavPath, Buffer.concat(this._audioBuffersForFile))
+          this._audioFilePath = wavPath
+          logger.info(`회의 오디오 저장: ${wavPath}`)
+        } catch (err) {
+          logger.warn(`오디오 저장 실패: ${err instanceof Error ? err.message : String(err)}`)
+        }
+      }
+      this._audioBuffersForFile = []
+
       // Step 3: 알림
       this._sendProgress(sessionId, 'notifying', 95)
       new Notification({
@@ -487,6 +511,35 @@ class MeetingModeService extends EventEmitter {
     this._segments = []
     this._memos = []
     this._meetingModeActive = false
+    this._audioBuffersForFile = []
+    this._audioFilePath = null
+  }
+
+  /** PCM16 버퍼를 WAV 파일로 저장 (16kHz mono 16-bit) */
+  private _saveWav(filePath: string, pcmBuffer: Buffer): void {
+    const sampleRate = 16000
+    const numChannels = 1
+    const bitsPerSample = 16
+    const byteRate = sampleRate * numChannels * (bitsPerSample / 8)
+    const blockAlign = numChannels * (bitsPerSample / 8)
+    const dataSize = pcmBuffer.length
+
+    const header = Buffer.alloc(44)
+    header.write('RIFF', 0)
+    header.writeUInt32LE(36 + dataSize, 4)
+    header.write('WAVE', 8)
+    header.write('fmt ', 12)
+    header.writeUInt32LE(16, 16)          // fmt chunk size
+    header.writeUInt16LE(1, 20)           // PCM
+    header.writeUInt16LE(numChannels, 22)
+    header.writeUInt32LE(sampleRate, 24)
+    header.writeUInt32LE(byteRate, 28)
+    header.writeUInt16LE(blockAlign, 32)
+    header.writeUInt16LE(bitsPerSample, 34)
+    header.write('data', 36)
+    header.writeUInt32LE(dataSize, 40)
+
+    fs.writeFileSync(filePath, Buffer.concat([header, pcmBuffer]))
   }
 
   private _sendToRenderer(channel: string, data: unknown): void {
@@ -941,18 +994,105 @@ ${transcript}`
     if (!row) throw new D3ROError(ErrorCode.MeetingSessionNotFound, `세션 없음: ${sessionId}`)
     if (!row.rawTranscript) throw new D3ROError(ErrorCode.DiarizationFailed, '전사 텍스트가 없습니다')
 
-    const hfToken = configGet('hfToken') as string
-    if (!hfToken) throw new D3ROError(ErrorCode.DiarizationTokenRequired, 'HuggingFace 토큰이 설정되지 않았습니다')
-
     this._sendToRenderer(IPC_CHANNELS.MEETING_MODE.DIARIZATION_PROGRESS, { sessionId, percent: 10 })
 
-    // 실제 오디오가 없으므로 LLM 기반 화자 추정 수행
+    // 오디오 파일 존재 확인
+    const { app } = await import('electron')
+    const path = await import('path')
+    const audioPath = path.join(app.getPath('userData'), 'meeting-audio', `${sessionId}.wav`)
+    const hasAudioFile = fs.existsSync(audioPath)
+
+    const hfToken = configGet('hfToken') as string
+
+    if (hasAudioFile && hfToken) {
+      // Phase 15.5-2: pyannote 실제 화자 구분
+      logger.info(`pyannote 화자 구분 시작: ${audioPath}`)
+      this._sendToRenderer(IPC_CHANNELS.MEETING_MODE.DIARIZATION_PROGRESS, { sessionId, percent: 20 })
+
+      try {
+        const audioBuffer = fs.readFileSync(audioPath)
+        const serverUrl = `http://127.0.0.1:18765`
+
+        const formData = new FormData()
+        formData.append('audio', new Blob([audioBuffer]), 'audio.wav')
+        formData.append('hf_token', hfToken)
+        if (numSpeakers && numSpeakers > 0) {
+          formData.append('num_speakers', String(numSpeakers))
+        }
+
+        this._sendToRenderer(IPC_CHANNELS.MEETING_MODE.DIARIZATION_PROGRESS, { sessionId, percent: 40 })
+
+        const response = await fetch(`${serverUrl}/diarize`, {
+          method: 'POST',
+          body: formData,
+        })
+
+        if (!response.ok) {
+          const errData = await response.json() as { message?: string }
+          throw new Error(errData.message ?? `Diarization failed: ${response.status}`)
+        }
+
+        const diarResult = await response.json() as {
+          segments: Array<{ speaker: string; start: number; end: number }>
+          num_speakers: number
+        }
+
+        this._sendToRenderer(IPC_CHANNELS.MEETING_MODE.DIARIZATION_PROGRESS, { sessionId, percent: 70 })
+
+        // 전사 세그먼트와 화자 세그먼트 매칭
+        const transcript = row.editedTranscript ?? row.rawTranscript ?? ''
+        const lines = transcript.split('\n').filter((l) => l.trim())
+        const diarSegments = diarResult.segments
+
+        const labeledLines = lines.map((line) => {
+          const timeMatch = /^\[(\d{2}):(\d{2})\]/.exec(line)
+          if (!timeMatch) return line
+
+          const timeSec = parseInt(timeMatch[1], 10) * 60 + parseInt(timeMatch[2], 10)
+          // 해당 시간에 가장 가까운 화자 세그먼트 찾기
+          const matchedSeg = diarSegments.find(
+            (s) => timeSec >= s.start && timeSec <= s.end,
+          )
+          if (matchedSeg) {
+            const speakerLabel = matchedSeg.speaker.replace('SPEAKER_', '화자 ')
+            return line.replace(/^\[(\d{2}:\d{2})\]/, `[$1] [${speakerLabel}]`)
+          }
+          return line
+        })
+
+        this._sendToRenderer(IPC_CHANNELS.MEETING_MODE.DIARIZATION_PROGRESS, { sessionId, percent: 85 })
+
+        db.update(meetingSessions).set({
+          editedTranscript: labeledLines.join('\n'),
+          updatedAt: Date.now(),
+        }).where(eq(meetingSessions.id, sessionId)).run()
+
+        logger.info(`pyannote 화자 구분 완료: ${diarResult.num_speakers}명, sessionId=${sessionId}`)
+      } catch (err) {
+        logger.warn(`pyannote 실패, LLM 폴백: ${err instanceof Error ? err.message : String(err)}`)
+        // pyannote 실패 시 LLM 폴백
+        await this._diarizeLLMFallback(sessionId, row, numSpeakers)
+      }
+    } else {
+      // LLM 기반 화자 추정 (오디오 없거나 HF 토큰 없음)
+      await this._diarizeLLMFallback(sessionId, row, numSpeakers)
+    }
+
+    this._sendToRenderer(IPC_CHANNELS.MEETING_MODE.DIARIZATION_PROGRESS, { sessionId, percent: 100 })
+    logger.info(`화자 구분 완료: sessionId=${sessionId}`)
+  }
+
+  private async _diarizeLLMFallback(
+    sessionId: string,
+    row: { rawTranscript: string | null; editedTranscript: string | null },
+    numSpeakers?: number,
+  ): Promise<void> {
     this._sendToRenderer(IPC_CHANNELS.MEETING_MODE.DIARIZATION_PROGRESS, { sessionId, percent: 30 })
 
     const { getLocalLLMService } = await import('./LocalLLMService')
     const llm = getLocalLLMService()
 
-    const transcript = row.editedTranscript ?? row.rawTranscript
+    const transcript = row.editedTranscript ?? row.rawTranscript ?? ''
     const speakerHint = numSpeakers && numSpeakers > 0
       ? `회의에는 총 ${numSpeakers}명의 화자가 있습니다.`
       : '화자 수는 문맥에서 추정하세요.'
@@ -970,13 +1110,11 @@ ${speakerHint}
 
     this._sendToRenderer(IPC_CHANNELS.MEETING_MODE.DIARIZATION_PROGRESS, { sessionId, percent: 80 })
 
+    const db = getDatabase()
     db.update(meetingSessions).set({
       editedTranscript: result.text,
       updatedAt: Date.now(),
     }).where(eq(meetingSessions.id, sessionId)).run()
-
-    this._sendToRenderer(IPC_CHANNELS.MEETING_MODE.DIARIZATION_PROGRESS, { sessionId, percent: 100 })
-    logger.info(`화자 구분 완료: sessionId=${sessionId}`)
   }
 
   private _buildPdfHtml(session: MeetingSessionDetail): string {