feat: 将视频转写重构为纯文字异步语音识别

app/api/admin/stt/videos/route.ts

@@ -1,26 +1,23 @@
 import { NextResponse } from "next/server";
-import { prisma } from "@/lib/prisma";
 import { getFileUrl } from "@/lib/minio";
+import { prisma } from "@/lib/prisma";
+import { transcriptionState } from "@/lib/stt/types";
 
 export async function GET() {
   try {
     const videos = await prisma.video.findMany({
       orderBy: { created_at: "desc" },
-      take: 1000,
       include: {
         author: {
           select: {
             nickname: true,
           },
         },
+        transcriptionTask: true,
         transcript: {
           select: {
             id: true,
-            speech_detected: true,
-            language: true,
-            audio_type: true,
             transcript: true,
-            non_speech_summary: true,
           },
         },
       },
@@ -29,21 +26,12 @@ export async function GET() {
     const formattedVideos = videos.map((v) => ({
       aweme_id: v.aweme_id,
       desc: v.desc,
-      cover_url: getFileUrl(v.cover_url ?? ""),
+      cover_url: v.cover_url ? getFileUrl(v.cover_url) : "",
       duration_ms: v.duration_ms,
       author: {
         nickname: v.author.nickname,
       },
-      transcript: v.transcript
-        ? {
-            id: v.transcript.id,
-            speech_detected: v.transcript.speech_detected,
-            language: v.transcript.language,
-            audio_type: v.transcript.audio_type,
-            transcript: v.transcript.transcript,
-            non_speech_summary: v.transcript.non_speech_summary,
-          }
-        : null,
+      transcription: transcriptionState(v.transcript, v.transcriptionTask),
     }));
 
     return NextResponse.json({

app/api/fetcher/index.ts

@@ -21,7 +21,7 @@ import {
   releaseIsolatedContext,
 } from "@/app/api/fetcher/browser";
 import { extractFirstFrame } from "@/app/api/media";
-import { transcriptAweme } from "../stt";
+import { enqueueTranscription } from "@/lib/stt/tasks";
 
 const DETAIL_PATH = "/aweme/v1/web/aweme/detail/";
 const COMMENT_PATH = "/aweme/v1/web/comment/list/";
@@ -429,7 +429,11 @@ export async function scrapeDouyin(url: string) {
         fps ?? undefined,
       );
       console.log(chalk.green.bold("✓ 视频作品保存成功"));
-      transcriptAweme(detail.aweme_detail.aweme_id).catch((e) => {}); // 异步转写，不阻塞主流程
+      try {
+        await enqueueTranscription(detail.aweme_detail.aweme_id);
+      } catch {
+        console.error("视频已保存，但转写任务入队失败，可在后台批量补转");
+      }
       return { type: "video", ...saved };
     } else {
       throw new ScrapeError(

app/api/media.ts

@@ -111,17 +111,17 @@ export async function getVideoDuration(
 }
 
 /**
- * 使用 ffmpeg 将视频二进制转换为音频，返回音频 Buffer
+ * 使用 ffmpeg 将视频 Buffer 或本地文件转换为音频，返回音频 Buffer
  *
  * 默认输出 mp3，可通过 opts.format 指定 'mp3' | 'aac' | 'wav'
- * - mp3: 使用 libmp3lame，contentType 为 'audio/mpeg'
+ * - mp3: 使用 libmp3lame，16 kHz 单声道，contentType 为 'audio/mpeg'
  * - aac: 输出 m4a 容器（AAC 编码），contentType 为 'audio/mp4'
  * - wav: PCM 16-bit，44100Hz，双声道，contentType 为 'audio/wav'
  *
  * 若系统未安装 ffmpeg 或未在 PATH 中，返回 null 并给出提示。
  */
 export async function extractAudio(
-  videoBuffer: Buffer,
+  videoBuffer: Buffer | string,
   opts?: { format?: "mp3" | "aac" | "wav"; bitrateKbps?: number },
 ): Promise<{ buffer: Buffer; contentType: string; ext: string } | null> {
   const ffmpegCmd = process.env.FFMPEG_PATH || "ffmpeg";
@@ -129,6 +129,7 @@ export async function extractAudio(
   const base = `dy_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
   const inPath = path.join(tmpDir, `${base}.mp4`);
 
+  const inputPath = typeof videoBuffer === "string" ? videoBuffer : inPath;
   const format = opts?.format ?? "mp3";
   const bitrate = Math.max(32, Math.min(512, opts?.bitrateKbps ?? 192)); // 安全范围 32~512 kbps
 
@@ -174,19 +175,28 @@ export async function extractAudio(
   }
 
   try {
-    await fs.writeFile(inPath, videoBuffer);
+    if (typeof videoBuffer !== "string") await fs.writeFile(inPath, videoBuffer);
     const args = [
       "-hide_banner",
       "-loglevel",
       "error",
       "-i",
-      inPath,
+      inputPath,
+      "-map",
+      "0:a:0",
+      "-ac",
+      "1",
+      "-ar",
+      "16000",
       "-vn", // 丢弃视频流
       ...codecArgs,
       "-y",
       outPath,
     ];
-    await execFileAsync(ffmpegCmd, args, { windowsHide: true });
+    await execFileAsync(ffmpegCmd, args, {
+      windowsHide: true,
+      timeout: 60 * 60 * 1000,
+    });
     const audio = await fs.readFile(outPath);
     return { buffer: audio, contentType, ext };
   } catch (e: any) {

app/api/stt/batch/route.ts

@@ -0,0 +1,22 @@
+import { NextResponse } from "next/server";
+import { getAsrConfig } from "@/lib/stt/dashscope";
+import { enqueueMissingTranscriptions } from "@/lib/stt/tasks";
+export const runtime = "nodejs";
+export async function POST() {
+  try {
+    getAsrConfig();
+  } catch (error) {
+    return NextResponse.json(
+      { error: (error as Error).message },
+      { status: 503 },
+    );
+  }
+  try {
+    return NextResponse.json(
+      { queued: await enqueueMissingTranscriptions() },
+      { status: 202 },
+    );
+  } catch {
+    return NextResponse.json({ error: "批量任务入队失败" }, { status: 500 });
+  }
+}

app/api/stt/index.ts

@@ -1,113 +0,0 @@
-import fs from "fs";
-import OpenAI from "openai";
-import prompt from "./prompt.md";
-import { prisma } from "@/lib/prisma";
-import { downloadFile } from "@/lib/minio";
-import { extractAudio } from "../media";
-import { z } from "zod";
-import { zodResponseFormat } from "openai/helpers/zod";
-import { ProxyAgent, setGlobalDispatcher } from "undici";
-
-const proxy = process.env.HTTPS_PROXY || process.env.HTTP_PROXY;
-if (proxy) {
-  setGlobalDispatcher(new ProxyAgent(proxy));
-}
-
-const client = new OpenAI({
-  apiKey: process.env.OPENAI_API_KEY,
-  baseURL: process.env.OPENAI_API_BASE_URL,
-});
-
-// Zod schema aligned with prompt.md
-export const SttSchema = z
-  .object({
-    speech_detected: z.boolean(),
-    language: z.string().min(1).nullable(),
-    audio_type: z.string().nullable(),
-    transcript: z.array(z.string()),
-    non_speech_summary: z.string().nullable(),
-  })
-  .strict();
-
-export type SttResult = z.infer<typeof SttSchema>;
-
-async function transcriptAudio(audio: Buffer | string) {
-  if (typeof audio === "string") {
-    audio = fs.readFileSync(audio);
-  }
-  const base64Audio = Buffer.from(audio).toString("base64");
-
-  const completion = await client.chat.completions.create({
-    model: "gemini-2.5-flash",
-    messages: [
-      {
-        role: "user",
-        content: [
-          { type: "text", text: prompt },
-          {
-            type: "input_audio",
-            input_audio: { data: base64Audio, format: "mp3" },
-          },
-        ],
-      },
-    ],
-    response_format: zodResponseFormat(SttSchema, "stt_result"),
-  });
-
-  const data = completion.choices?.[0]?.message;
-  console.log("转写结果", data.content);
-
-  if (!data || !data.content) {
-    throw new Error("No STT result returned from model");
-  }
-  try {
-    return JSON.parse(data.content) as SttResult;
-  } catch (e) {
-    throw new Error("Failed to parse STT result JSON");
-  }
-}
-
-export async function transcriptAweme(awemeId: string): Promise<SttResult> {
-  const aweme = await prisma.video.findUnique({
-    where: { aweme_id: awemeId },
-  });
-  if (!aweme) {
-    throw new Error("Aweme not found or aweme is not a video post");
-  }
-  const vPath = aweme.video_url;
-  const buffer = await downloadFile(vPath);
-
-  const audioDat = await extractAudio(buffer, {
-    format: "mp3",
-    bitrateKbps: 128,
-  });
-
-  if (!audioDat || !audioDat.buffer) {
-    throw new Error("Failed to extract audio from video");
-  }
-
-  // 调用大模型生成转写结果
-  const result = await transcriptAudio(audioDat.buffer);
-
-  // 将转写结果持久化到数据库
-  await prisma.videoTranscript.upsert({
-    where: { videoId: awemeId },
-    update: {
-      speech_detected: result.speech_detected,
-      language: result.language,
-      audio_type: result.audio_type,
-      transcript: result.transcript,
-      non_speech_summary: result.non_speech_summary,
-    },
-    create: {
-      videoId: awemeId,
-      speech_detected: result.speech_detected,
-      language: result.language,
-      audio_type: result.audio_type,
-      transcript: result.transcript,
-      non_speech_summary: result.non_speech_summary,
-    },
-  });
-
-  return result;
-}

app/api/stt/prompt.md

@@ -1,29 +0,0 @@
-你将接收一段音频。请完成：你将接收一段音频。语音活动检测（VAD）与声源分类。
-
-**输出 JSON Schema（示例）**
-
-```json
-{
-  "speech_detected": true,
-  "language": "zh-CN",
-  "audio_type": null,
-  "transcript": ["大家好，我是xxx", "欢迎来到今天的视频", "今天我们来聊聊AI的未来"],
-  "non_speech_summary": null,
-}
-```
-
-**当无发言时返回：** 
-
-```json
-{
-  "speech_detected": false,
-  "language": null,
-  "audio_type": "music | ambience | animal | mechanical | other",
-  "transcript": [],
-  "non_speech_summary": "示例：纯音乐-钢琴独奏，节奏舒缓；或 环境声-雨声伴随雷鸣。",
-}
-```
-
-** 规则补充 **
-
-* 允许`language` 为多标签（如 "zh-CN,en"）或为 `null`（无发言时）。

app/api/stt/route.ts

@@ -1,23 +1,33 @@
+import { type NextRequest, NextResponse } from "next/server";
+import { enqueueTranscription, readTranscription } from "@/lib/stt/tasks";
+import { isActive } from "@/lib/stt/types";
 export const runtime = "nodejs";
-import { NextRequest, NextResponse } from "next/server";
-import { prisma } from "@/lib/prisma";
-import type { FeedItem, FeedResponse } from "@/app/types/feed";
-import { getFileUrl } from "@/lib/minio";
-import { transcriptAweme } from ".";
-
-// Contract
-// Inputs: search params { before?: ISOString, limit?: number }
-// Output: { items: FeedItem[], nextCursor: ISOString | null }
-
+export const dynamic = "force-dynamic";
+const validId = (id: unknown): id is string =>
+  typeof id === "string" && /^[A-Za-z0-9_-]{1,128}$/.test(id);
 export async function GET(req: NextRequest) {
-  const { searchParams } = new URL(req.url);
-  const awemeId = searchParams.get("awemeId");
-  if (!awemeId) {
-    return NextResponse.json(
-      { error: "Missing awemeId parameter" },
-      { status: 400 },
-    );
+  const id = req.nextUrl.searchParams.get("awemeId");
+  if (!validId(id))
+    return NextResponse.json({ error: "需要有效的 awemeId" }, { status: 400 });
+  try {
+    const state = await readTranscription(id);
+    return state
+      ? NextResponse.json(state)
+      : NextResponse.json({ error: "视频不存在" }, { status: 404 });
+  } catch {
+    return NextResponse.json({ error: "读取转写状态失败" }, { status: 500 });
+  }
+}
+export async function POST(req: NextRequest) {
+  const body = await req.json().catch(() => null);
+  if (!validId(body?.awemeId))
+    return NextResponse.json({ error: "需要有效的 awemeId" }, { status: 400 });
+  try {
+    const state = await enqueueTranscription(body.awemeId, true);
+    return state
+      ? NextResponse.json(state, { status: isActive(state.status) ? 202 : 200 })
+      : NextResponse.json({ error: "视频不存在" }, { status: 404 });
+  } catch {
+    return NextResponse.json({ error: "转写任务入队失败" }, { status: 500 });
   }
-  const script = await transcriptAweme(awemeId);
-  return NextResponse.json(script);
 }