auto-save 2026-05-14 11:21 (~7)

api/main.py

@@ -33,6 +33,7 @@ PRODUCT_LIBRARY_MANIFEST = PRODUCT_LIBRARY_DIR / "manifest.json"
 LLM_BASE_URL = os.getenv("LLM_BASE_URL", "").strip()
 LLM_API_KEY = os.getenv("LLM_API_KEY", "").strip()
 ASR_MODEL = os.getenv("ASR_MODEL", "whisper-1")
+ASR_FALLBACK_MODEL = os.getenv("ASR_FALLBACK_MODEL", "gemini-2.5-flash").strip() or "gemini-2.5-flash"
 TRANSLATE_MODEL = os.getenv("TRANSLATE_MODEL", "gemini-2.5-flash")
 REWRITE_MODEL = os.getenv("REWRITE_MODEL", "gemini-2.5-pro")
 VISION_MODEL = os.getenv("VISION_MODEL", "gemini-2.5-flash")
@@ -687,8 +688,8 @@ def _resolve_frame_quality(duration: float, quality: FrameExtractQuality) -> Fra
     cores = os.cpu_count() or 4
     memory_gb = _physical_memory_gb()
     strong_machine = cores >= 10 and (memory_gb == 0.0 or memory_gb >= 32)
-    if strong_machine and duration <= 180:
-        return "ultra"
+    # 展示/演示时不能把本机资源打满：auto 最高只到 accurate。
+    # ultra 保留为手动选择项，不再由 auto 自动命中。
     if strong_machine and duration <= 600:
         return "accurate"
     if cores >= 8 and duration <= 240:
@@ -1157,6 +1158,16 @@ def ffprobe_meta(mp4: Path) -> dict:
     return json.loads(out)
 
 
+def media_duration(path: Path) -> float:
+    try:
+        out = run([
+            "ffprobe", "-v", "error", "-print_format", "json", "-show_format", str(path),
+        ])
+        return float(json.loads(out).get("format", {}).get("duration") or 0)
+    except Exception:
+        return 0.0
+
+
 def pipeline_download(job_id: str) -> None:
     """阶段 1：仅下载（或上传跳过），落 source.mp4，停在 downloaded 等用户点解析/提取音频。"""
     job = JOBS[job_id]
@@ -1362,21 +1373,83 @@ def analyze_queue_worker() -> None:
 
 # ---------- 音频转写 + 翻译 + SKG 改写 + MiniMax 配音 ----------
 
+def _parse_asr_segments(content: str, duration: float) -> list[dict]:
+    raw = (content or "").strip()
+    if raw.startswith("```"):
+        import re as _re
+        match = _re.search(r"(\[[\s\S]*\]|\{[\s\S]*\})", raw)
+        raw = match.group(0) if match else raw
+    try:
+        data = json.loads(raw)
+    except json.JSONDecodeError:
+        text = raw.strip()
+        return [{"start": 0.0, "end": duration, "text": text}] if text else []
+    if isinstance(data, dict):
+        for key in ("segments", "data", "items", "result"):
+            if isinstance(data.get(key), list):
+                data = data[key]
+                break
+        else:
+            text = str(data.get("text") or data.get("transcript") or "").strip()
+            return [{"start": 0.0, "end": duration, "text": text}] if text else []
+    if not isinstance(data, list):
+        return []
+    segments: list[dict] = []
+    for i, item in enumerate(data):
+        if isinstance(item, str):
+            text = item.strip()
+            start = 0.0 if len(data) == 1 else duration * i / max(1, len(data))
+            end = duration if len(data) == 1 else duration * (i + 1) / max(1, len(data))
+        elif isinstance(item, dict):
+            text = str(item.get("text") or item.get("en") or item.get("transcript") or "").strip()
+            start = float(item.get("start") or item.get("start_time") or 0)
+            end = float(item.get("end") or item.get("end_time") or duration)
+        else:
+            continue
+        if text:
+            segments.append({"start": max(0.0, start), "end": max(start, end), "text": text})
+    return segments
+
+
+def _transcribe_gemini_sync(wav: Path) -> list[dict]:
+    duration = media_duration(wav)
+    audio_b64 = base64.b64encode(wav.read_bytes()).decode("ascii")
+    prompt = (
+        "Transcribe the attached audio. Return strict JSON only, no markdown. "
+        "Schema: [{\"start\": 0.0, \"end\": 1.2, \"text\": \"English transcript\"}]. "
+        "Use English for the transcript. If exact timestamps are uncertain, return one segment "
+        f"from 0 to {duration:.2f} seconds."
+    )
+    resp = llm().chat.completions.create(
+        model=ASR_FALLBACK_MODEL,
+        messages=[{"role": "user", "content": [
+            {"type": "text", "text": prompt},
+            {"type": "input_audio", "input_audio": {"data": audio_b64, "format": "wav"}},
+        ]}],
+        temperature=0,
+    )
+    content = (resp.choices[0].message.content or "").strip()
+    return _parse_asr_segments(content, duration)
+
+
 def _transcribe_sync(wav: Path) -> list[dict]:
     """whisper-1 verbose_json → segments[{start, end, text}]"""
-    with wav.open("rb") as f:
-        resp = llm().audio.transcriptions.create(
-            file=(wav.name, f, "audio/wav"),
-            model=ASR_MODEL,
-            response_format="verbose_json",
-            timestamp_granularities=["segment"],
-        )
-    raw = resp.model_dump() if hasattr(resp, "model_dump") else resp
-    segments = raw.get("segments") or []
-    # 兜底：网关如果不返回 segments，把全文当一段
-    if not segments and raw.get("text"):
-        segments = [{"start": 0.0, "end": float(raw.get("duration", 0) or 0), "text": raw["text"]}]
-    return segments
+    try:
+        with wav.open("rb") as f:
+            resp = llm().audio.transcriptions.create(
+                file=(wav.name, f, "audio/wav"),
+                model=ASR_MODEL,
+                response_format="verbose_json",
+                timestamp_granularities=["segment"],
+            )
+        raw = resp.model_dump() if hasattr(resp, "model_dump") else resp
+        segments = raw.get("segments") or []
+        # 兜底：网关如果不返回 segments，把全文当一段
+        if not segments and raw.get("text"):
+            segments = [{"start": 0.0, "end": float(raw.get("duration", 0) or 0), "text": raw["text"]}]
+        return segments
+    except Exception:
+        return _transcribe_gemini_sync(wav)
 
 
 def _translate_sync(segments: list[dict]) -> list[str]:
@@ -1865,6 +1938,7 @@ def health() -> dict:
         "base_url": LLM_BASE_URL or "openai-default",
         "models": {
             "asr": ASR_MODEL,
+            "asr_fallback": ASR_FALLBACK_MODEL,
             "translate": TRANSLATE_MODEL,
             "rewrite": REWRITE_MODEL,
             "audio_rewrite": AUDIO_REWRITE_MODEL,