auto-save 2026-05-14 11:47 (~7)

api/.env.example

@@ -19,7 +19,7 @@ AUDIO_PRODUCT_BRIEF="SKG 智能按摩产品，主打日常肩颈、腰背、眼
 MINIMAX_API_KEY=
 MINIMAX_TTS_BASE_URL=https://api.minimax.io
 MINIMAX_TTS_MODEL=speech-2.8-turbo
-MINIMAX_TTS_VOICE_ID="Chinese (Mandarin)_Reliable_Executive"
+MINIMAX_TTS_VOICE_ID=English_expressive_narrator
 
 # Poe 视频 API（优先用于 Seedance / Kling / Veo）
 POE_API_BASE_URL=https://api.poe.com/v1

api/main.py

@@ -49,8 +49,8 @@ MINIMAX_TTS_BASE_URL = os.getenv("MINIMAX_TTS_BASE_URL", "https://api.minimax.io
 MINIMAX_TTS_MODEL = os.getenv("MINIMAX_TTS_MODEL", "speech-2.8-turbo").strip() or "speech-2.8-turbo"
 MINIMAX_TTS_VOICE_ID = os.getenv(
     "MINIMAX_TTS_VOICE_ID",
-    "Chinese (Mandarin)_Reliable_Executive",
-).strip() or "Chinese (Mandarin)_Reliable_Executive"
+    "English_expressive_narrator",
+).strip() or "English_expressive_narrator"
 
 POE_API_BASE_URL = os.getenv("POE_API_BASE_URL", "https://api.poe.com/v1").strip() or "https://api.poe.com/v1"
 POE_API_KEY = os.getenv("POE_API_KEY", "").strip()
@@ -379,6 +379,7 @@ class Job(BaseModel):
     duration: float = 0.0
     width: int = 0
     height: int = 0
+    source_audio_url: str = ""
     frames: list[KeyFrame] = Field(default_factory=list)
     transcript: list[TranscriptSegment] = Field(default_factory=list)
     audio_script: AudioScript = Field(default_factory=AudioScript)
@@ -400,6 +401,14 @@ def job_dir(job_id: str) -> Path:
     return d
 
 
+def source_audio_url_for(job_id: str) -> str:
+    return f"/jobs/{job_id}/audio.wav" if (JOBS_DIR / job_id / "audio.wav").exists() else ""
+
+
+def job_with_artifacts(job: Job) -> Job:
+    return job.model_copy(update={"source_audio_url": source_audio_url_for(job.id)})
+
+
 def save_state(job: Job) -> None:
     (job_dir(job.id) / "state.json").write_text(job.model_dump_json(indent=2))
 
@@ -1224,7 +1233,7 @@ def pipeline_analyze(
 
         wav = d / "audio.wav"
         if wav.exists():
-            update(job, status="splitting", message="复用音轨 · 准备抽帧…", progress=35)
+            update(job, status="splitting", message="复用音轨 · 准备抽帧…", progress=35, source_audio_url=f"/jobs/{job_id}/audio.wav")
         else:
             update(job, status="splitting", message="ffmpeg 拆分音轨…", progress=35)
             run([
@@ -1232,6 +1241,7 @@ def pipeline_analyze(
                 "-vn", "-ac", "1", "-ar", "16000", "-c:a", "pcm_s16le",
                 str(wav),
             ])
+            update(job, source_audio_url=f"/jobs/{job_id}/audio.wav")
         n = max(1, min(int(frame_count), 20))
         target_label = FRAME_TARGET_LABELS.get(target, FRAME_TARGET_LABELS["balanced"])
         duration = max(float(job.duration or 1.0), 0.1)
@@ -1497,12 +1507,12 @@ def _transcript_join(segments: list[TranscriptSegment], field: Literal["en", "zh
 
 
 def _fallback_audio_script(segments: list[TranscriptSegment]) -> str:
-    joined = " ".join((s.zh or s.en).strip() for s in segments if (s.zh or s.en).strip())
+    joined = " ".join((s.en or s.zh).strip() for s in segments if (s.en or s.zh).strip())
     if not joined:
-        return "日常疲惫不用硬扛。戴上 SKG，让肩颈慢慢放松，跟着呼吸找回轻松状态。"
+        return "Ease into the moment with SKG. Gentle warmth and rhythmic massage help everyday tension feel lighter, cleaner, and easier to leave behind."
     return (
-        "把日常紧绷交给 SKG。贴合身体需要放松的位置，热敷与按摩节奏自然陪伴，"
-        "让每一次短暂休息都更轻松、更有质感。"
+        "Let SKG turn a short break into real relief. With soothing warmth and steady massage rhythm, "
+        "everyday tension feels lighter, calmer, and easier to leave behind."
     )
 
 
@@ -1513,24 +1523,24 @@ def _rewrite_audio_script_sync(segments: list[TranscriptSegment]) -> tuple[str,
     source_text = _transcript_join(segments, "en")
     source_zh = _transcript_join(segments, "zh")
     prompt = (
-        "你是 SKG 短视频口播编导。根据参考视频音频转写，抽取它的表达结构、情绪节奏和可复用卖点，"
-        "改写成适合 SKG 按摩/放松产品二创视频的中文口播文案。\n"
-        "要求：\n"
-        "1. 输出 35-90 个中文字，适合 8-18 秒短视频配音。\n"
-        "2. 口语化、干净、高级，能直接给 TTS 朗读。\n"
-        "3. 不承诺治疗、治愈、医学疗效，不夸大。\n"
-        "4. 不复刻原视频品牌/人物/价格/平台话术，只保留表达结构。\n"
-        "5. 如果参考转写信息不足，按产品信息生成通用 SKG 放松口播。\n"
-        '严格返回 JSON：{"rewritten_text":"..."}。\n\n'
-        f"SKG 产品信息：{AUDIO_PRODUCT_BRIEF}\n\n"
-        f"英文转写：\n{source_text or '无'}\n\n"
-        f"中文翻译：\n{source_zh or '无'}"
+        "You are an English short-video voice-over writer for SKG wellness massagers. "
+        "Use the source transcript only for structure, pacing, and emotional hook, then rewrite it into a clean English VO for SKG.\n"
+        "Rules:\n"
+        "1. Output 28-55 English words, suitable for an 8-18 second TTS voice-over.\n"
+        "2. Make it natural, premium, concise, and ready to read aloud.\n"
+        "3. Do not claim medical treatment, cure, pain elimination, or clinical effects.\n"
+        "4. Do not copy the original brand, creator, price, platform language, or exact claims.\n"
+        "5. If the source transcript is too thin, write a general SKG relaxation VO.\n"
+        'Return strict JSON only: {"rewritten_text":"..."}.\n\n'
+        f"SKG product context: {AUDIO_PRODUCT_BRIEF}\n\n"
+        f"English transcript:\n{source_text or 'None'}\n\n"
+        f"Chinese translation for reference:\n{source_zh or 'None'}"
     )
     try:
         resp = llm().chat.completions.create(
             model=AUDIO_REWRITE_MODEL,
             messages=[
-                {"role": "system", "content": "只输出合法 JSON，不要解释，不要 markdown。"},
+                {"role": "system", "content": "Return valid JSON only. No explanation. No markdown."},
                 {"role": "user", "content": prompt},
             ],
             response_format={"type": "json_object"},
@@ -1564,7 +1574,7 @@ def _minimax_tts_sync(job_id: str, text: str) -> str:
         "model": MINIMAX_TTS_MODEL,
         "text": text.strip()[:9500],
         "stream": False,
-        "language_boost": "Chinese",
+        "language_boost": "English",
         "output_format": "hex",
         "voice_setting": {
             "voice_id": MINIMAX_TTS_VOICE_ID,
@@ -1651,6 +1661,7 @@ def pipeline_transcribe(job_id: str, manage_job_status: bool = True) -> None:
             ])
             if not wav.exists():
                 raise RuntimeError("音频提取完成但找不到 audio.wav")
+        update(job, source_audio_url=f"/jobs/{job_id}/audio.wav")
 
         if not LLM_API_KEY:
             # 无 key 模式：mock 数据
@@ -2112,7 +2123,7 @@ def get_job(job_id: str) -> Job:
     job = JOBS.get(job_id)
     if not job:
         raise HTTPException(404, "job not found")
-    return job
+    return job_with_artifacts(job)
 
 
 @app.delete("/jobs/{job_id}")
@@ -2153,7 +2164,7 @@ async def trigger_transcribe(job_id: str, bg: BackgroundTasks) -> Job:
         update(job, error="", audio_script=audio_payload)
     if not start_audio_processing(job_id, manage_job_status=manage_job_status):
         update(job, message="音频已在处理中")
-    return job
+    return job_with_artifacts(job)
 
 
 @app.get("/jobs/{job_id}/video.mp4")