feat: 實作 Groq Whisper 影片轉錄（Issue #4）

Dockerfile

@@ -1,5 +1,11 @@
 FROM node:20-slim
 
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    python3 ffmpeg curl \
+    && curl -L https://github.com/yt-dlp/yt-dlp/releases/latest/download/yt-dlp -o /usr/local/bin/yt-dlp \
+    && chmod a+rx /usr/local/bin/yt-dlp \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
 WORKDIR /app
 
 COPY package.json package-lock.json* ./

README.md

@@ -56,6 +56,10 @@ LLM_API_KEY=...
 LLM_MODEL=MiniMaxAI/MiniMax-M2.5
 TG_BOT_TOKEN=...
 TG_CHANNEL_ID=-100...
+
+# 影片轉錄（選填，啟用後自動轉錄影片貼文）
+TRANSCRIBER=groq
+GROQ_API_KEY=gsk_...
 ```
 
 ## CLI 工具模式
@@ -123,6 +127,7 @@ npx @cablate/banini-tracker push -m "分析結果..."
 |------|---------|------|--------|
 | Facebook 抓取（Apify） | ~$0.02 | 盤中 ~198 次 + 盤後 30 次 | ~$4.56 |
 | LLM 分析（常駐模式） | 依模型而定 | 同上 | 依模型定價 |
+| 影片轉錄（Groq Whisper） | ~$0.006/分鐘 | 視影片數量 | 極低 |
 | Telegram 推送 | 免費 | — | $0 |
 
 > 盤中：週一~五 09:00-13:30 每 30 分鐘（~9 次/日 × 22 工作日）

package.json

@@ -30,6 +30,7 @@
   "dependencies": {
     "commander": "^13.0.0",
     "dotenv": "^16.4.0",
+    "groq-sdk": "^1.1.2",
     "node-cron": "^4.2.1",
     "openai": "^4.0.0"
   },

src/facebook.ts

@@ -57,7 +57,9 @@ export async function fetchFacebookPosts(
       shareCount: item.shares ?? 0,
       url: item.url ?? '',
       mediaType: media?.__typename?.toLowerCase() ?? 'text',
-      mediaUrl: media?.thumbnail ?? media?.photo_image?.uri ?? '',
+      mediaUrl: media?.video_url ?? media?.playable_url
+        ?? (media?.__typename?.toLowerCase() === 'video' ? media?.url : null)
+        ?? media?.thumbnail ?? media?.photo_image?.uri ?? '',
     };
   });
 }

src/index.ts

@@ -16,6 +16,7 @@ import { analyzePosts } from './analyze.js';
 import { sendTelegramMessageWithConfig, formatReport, formatFallbackReport } from './telegram.js';
 import { filterNewPosts as filterNew, markPostsSeen } from './seen.js';
 import { withRetry } from './retry.js';
+import { createTranscriber, transcribeVideoPosts, type TranscriberType } from './transcribe.js';
 
 // ── Config ──────────────────────────────────────────────────
 const FB_PAGE_URL = 'https://www.facebook.com/DieWithoutBang/';
@@ -41,6 +42,7 @@ interface UnifiedPost {
   mediaType: string;
   mediaUrl: string;
   ocrText: string;
+  transcriptText: string;
 }
 
 function fromFacebook(p: FacebookPost): UnifiedPost {
@@ -49,6 +51,7 @@ function fromFacebook(p: FacebookPost): UnifiedPost {
     source: 'facebook',
     text: p.text,
     ocrText: p.ocrText,
+    transcriptText: '',
     timestamp: p.timestamp,
     likeCount: p.likeCount,
     replyCount: p.commentCount,
@@ -112,6 +115,17 @@ async function runInner(opts: RunOptions) {
     return;
   }
 
+  // 2.5. 影片轉錄
+  const transcriberType = (process.env.TRANSCRIBER ?? 'noop') as TranscriberType;
+  const transcriber = createTranscriber(transcriberType);
+  if (transcriber.name !== 'noop') {
+    const transcripts = await transcribeVideoPosts(newPosts, transcriber);
+    for (const p of newPosts) {
+      const result = transcripts.get(p.id);
+      if (result) p.transcriptText = result.text;
+    }
+  }
+
   // 按時間從新到舊排序
   newPosts.sort((a, b) => new Date(b.timestamp).getTime() - new Date(a.timestamp).getTime());
 
@@ -135,6 +149,7 @@ async function runInner(opts: RunOptions) {
     const localTime = new Date(p.timestamp).toLocaleString('zh-TW', { timeZone: 'Asia/Taipei' });
     console.log(`--- [${tag}]${todayTag} ${localTime} [${p.mediaType}] ---`);
     console.log(p.text || '（無文字，可能是純圖片）');
+    if (p.transcriptText) console.log(`[影片轉錄] ${p.transcriptText}`);
     if (p.mediaUrl) console.log(`媒體: ${p.mediaUrl}`);
     console.log(`讚: ${p.likeCount} | 回覆: ${p.replyCount} | ${p.url}\n`);
   }
@@ -146,12 +161,13 @@ async function runInner(opts: RunOptions) {
 
   // 5. AI 分析
   const textsForAnalysis = newPosts
-    .filter((p) => p.text.trim().length > 0 || p.ocrText.trim().length > 0)
+    .filter((p) => p.text.trim().length > 0 || p.ocrText.trim().length > 0 || p.transcriptText.trim().length > 0)
     .map((p) => {
       const tag = 'Facebook';
       const localTime = new Date(p.timestamp).toLocaleString('zh-TW', { timeZone: 'Asia/Taipei' });
       let content = `[${tag}] ${p.text}`;
       if (p.ocrText) content += `\n[圖片 OCR] ${p.ocrText}`;
+      if (p.transcriptText) content += `\n[影片轉錄] ${p.transcriptText}`;
       return { text: content, timestamp: localTime, isToday: isToday(p.timestamp) };
     });
 

src/transcribe.ts

@@ -0,0 +1,171 @@
+// ── 影片轉錄抽象層 ──────────────────────────────────────────
+// 策略模式：定義轉錄介面，具體實作由外部決定。
+// 新增轉錄服務時只需實作 Transcriber 介面並在 createTranscriber() 加入。
+
+import { execFile } from 'child_process';
+import { createReadStream, unlinkSync, mkdirSync } from 'fs';
+import { join } from 'path';
+import { tmpdir } from 'os';
+import { promisify } from 'util';
+import Groq from 'groq-sdk';
+
+const execFileAsync = promisify(execFile);
+
+export interface TranscribeResult {
+  text: string;
+  durationSec?: number;
+}
+
+export interface Transcriber {
+  readonly name: string;
+  transcribe(videoUrl: string): Promise<TranscribeResult>;
+}
+
+// ── Noop（預設：不轉錄）────────────────────────────────────
+
+export class NoopTranscriber implements Transcriber {
+  readonly name = 'noop';
+  async transcribe(_videoUrl: string): Promise<TranscribeResult> {
+    return { text: '' };
+  }
+}
+
+// ── Groq Whisper ────────────────────────────────────────────
+// Facebook 影片 URL 通常是頁面連結（reel/watch），Groq 無法直接存取。
+// 流程：yt-dlp 下載音訊 → 傳檔案給 Groq Whisper → 清理暫存檔。
+
+function needsDownload(url: string): boolean {
+  return /facebook\.com\/(reel|watch|video)/i.test(url);
+}
+
+async function downloadAudio(videoUrl: string): Promise<string> {
+  const tmpDir = join(tmpdir(), 'banini-tracker');
+  mkdirSync(tmpDir, { recursive: true });
+  const outTemplate = join(tmpDir, `audio-${Date.now}.%(ext)s`);
+  const outFile = join(tmpDir, `audio-${Date.now()}.m4a`);
+
+  await execFileAsync('yt-dlp', [
+    '-f', 'ba',
+    '--extract-audio',
+    '--audio-format', 'm4a',
+    '-o', outFile.replace('.m4a', '.%(ext)s'),
+    '--no-playlist',
+    '--quiet',
+    videoUrl,
+  ], { timeout: 60_000 });
+
+  return outFile;
+}
+
+export class GroqTranscriber implements Transcriber {
+  readonly name = 'groq';
+  private client: Groq;
+  private model: string;
+
+  constructor(apiKey: string, model = 'whisper-large-v3') {
+    this.client = new Groq({ apiKey });
+    this.model = model;
+  }
+
+  async transcribe(videoUrl: string): Promise<TranscribeResult> {
+    if (needsDownload(videoUrl)) {
+      return this.transcribeViaDownload(videoUrl);
+    }
+    return this.transcribeViaUrl(videoUrl);
+  }
+
+  private async transcribeViaUrl(url: string): Promise<TranscribeResult> {
+    const result = await this.client.audio.transcriptions.create({
+      url,
+      model: this.model,
+      language: 'zh',
+      temperature: 0,
+      response_format: 'verbose_json',
+    });
+    const raw = result as any;
+    return {
+      text: result.text ?? '',
+      durationSec: raw.duration ? Math.round(raw.duration) : undefined,
+    };
+  }
+
+  private async transcribeViaDownload(videoUrl: string): Promise<TranscribeResult> {
+    console.log(`[轉錄] 下載音訊: ${videoUrl.slice(0, 60)}...`);
+    const audioFile = await downloadAudio(videoUrl);
+    try {
+      const result = await this.client.audio.transcriptions.create({
+        file: createReadStream(audioFile),
+        model: this.model,
+        language: 'zh',
+        temperature: 0,
+        response_format: 'verbose_json',
+      });
+      const raw = result as any;
+      return {
+        text: result.text ?? '',
+        durationSec: raw.duration ? Math.round(raw.duration) : undefined,
+      };
+    } finally {
+      try { unlinkSync(audioFile); } catch {}
+    }
+  }
+}
+
+// ── 工廠 ──────────────────────────────────────────────────
+
+export type TranscriberType = 'noop' | 'groq';
+
+export function createTranscriber(type: TranscriberType = 'noop'): Transcriber {
+  switch (type) {
+    case 'noop':
+      return new NoopTranscriber();
+    case 'groq': {
+      const apiKey = process.env.GROQ_API_KEY;
+      if (!apiKey) throw new Error('GROQ_API_KEY 環境變數未設定');
+      return new GroqTranscriber(apiKey, process.env.GROQ_WHISPER_MODEL ?? 'whisper-large-v3');
+    }
+    default:
+      throw new Error(`不支援的轉錄器類型: ${type}`);
+  }
+}
+
+// ── 輔助：判斷貼文是否為影片 ────────────────────────────────
+
+export function isVideoPost(mediaType: string): boolean {
+  const videoTypes = ['video', 'native_video', 'live_video', 'reel'];
+  return videoTypes.includes(mediaType.toLowerCase());
+}
+
+// ── 批次轉錄 ────────────────────────────────────────────────
+
+export interface TranscribablePost {
+  id: string;
+  mediaType: string;
+  mediaUrl: string;
+}
+
+export async function transcribeVideoPosts<T extends TranscribablePost>(
+  posts: T[],
+  transcriber: Transcriber,
+): Promise<Map<string, TranscribeResult>> {
+  const results = new Map<string, TranscribeResult>();
+
+  for (const post of posts) {
+    if (!isVideoPost(post.mediaType) || !post.mediaUrl) continue;
+
+    try {
+      console.log(`[轉錄][${transcriber.name}] 處理影片: ${post.id}`);
+      const result = await transcriber.transcribe(post.mediaUrl);
+      if (result.text.trim().length > 0) {
+        results.set(post.id, result);
+        console.log(`[轉錄] ${post.id}: ${result.text.slice(0, 50)}...（${result.durationSec ?? '?'}s）`);
+      } else {
+        console.log(`[轉錄] ${post.id}: 無可辨識內容`);
+      }
+    } catch (err) {
+      console.error(`[轉錄] ${post.id} 失敗: ${err instanceof Error ? err.message : err}`);
+    }
+  }
+
+  return results;
+}