feat: media API (video/bbox/thumbnail), UUID unification, dot matrix text, portal fixes, API dictionary V1.3

scripts/asr_processor.py

@@ -18,6 +18,7 @@ Configuration:
 import sys
 import json
 import os
+import time
 import argparse
 import signal
 import subprocess
@@ -179,7 +180,11 @@ def run_asr(video_path, output_path, uuid: str = ""):
     if publisher:
         publisher.info("asr", "Loading Whisper model...")
 
+    sys.stderr.write(f"[ASR] Loading Whisper model {MODEL_SIZE}...\n")
+    sys.stderr.flush()
     model = WhisperModel(MODEL_SIZE, device="cpu", compute_type="int8")
+    sys.stderr.write(f"[ASR] Model loaded\n")
+    sys.stderr.flush()
 
     if publisher:
         publisher.info("asr", f"Transcribing: {video_path}")
@@ -189,10 +194,13 @@ def run_asr(video_path, output_path, uuid: str = ""):
 
     if cut_scenes:
         # 分段處理：對每個場景萃取音訊並轉錄
+        sys.stderr.write(f"[ASR] Starting segmented transcription for {len(cut_scenes)} scenes\n")
+        sys.stderr.flush()
         import subprocess
         import tempfile
-        import json
         temp_dir = tempfile.mkdtemp(prefix="asr_cut_")
+        sys.stderr.write(f"[ASR] Temp dir: {temp_dir}\n")
+        sys.stderr.flush()
         transcript_language = None
 
         # 建立 scene lookup: 給定時間點，找是哪個 scene
@@ -204,29 +212,79 @@ def run_asr(video_path, output_path, uuid: str = ""):
 
         # 逐段處理，每段結果即時寫入 .asr.tmp
         tmp_path = output_path + ".tmp"
+        err_path = output_path + ".err"
         all_segments = []
 
+        # Resume: 若 executor 將 .tmp rename 成 .err，先救回
+        if not os.path.exists(tmp_path) and os.path.exists(err_path) and os.path.getsize(err_path) > 10:
+            try:
+                os.rename(err_path, tmp_path)
+                sys.stderr.write(f"[ASR] Recovered .err → .tmp for resume ({os.path.getsize(tmp_path)} bytes)\n")
+                sys.stderr.flush()
+            except Exception as e:
+                sys.stderr.write(f"[ASR] Failed to recover .err: {e}\n")
+                sys.stderr.flush()
+
+        # Resume: 若已有 .asr.tmp，載入已完成的 segments 並跳過已處理的 scenes
+        resume_from_scene = 0
+        if os.path.exists(tmp_path) and os.path.getsize(tmp_path) > 10:
+            try:
+                with open(tmp_path) as f:
+                    existing = json.load(f)
+                all_segments = existing.get("segments", [])
+                if all_segments:
+                    # 找出最後一個 segment 的 end_time，決定 resume 起點
+                    last_end = max(s.get("end", 0) for s in all_segments)
+                    # 找出最後完成的 scene_idx（場景 end_time > last_end）
+                    for i, (st, et) in enumerate(cut_scenes):
+                        if et > last_end:
+                            resume_from_scene = i
+                            break
+                    else:
+                        resume_from_scene = len(cut_scenes)  # 全部完成
+                    # 繼承 language
+                    if existing.get("language"):
+                        transcript_language = existing["language"]
+                    sys.stderr.write(f"[ASR] Resume from scene {resume_from_scene}/{len(cut_scenes)} "
+                                     f"(last segment end={last_end:.1f}s, {len(all_segments)} existing segments)\n")
+                    sys.stderr.flush()
+            except Exception as e:
+                sys.stderr.write(f"[ASR] Failed to load tmp for resume: {e}, starting fresh\n")
+                sys.stderr.flush()
+                all_segments = []
+
         for idx, (start_t, end_t) in enumerate(cut_scenes):
+            if idx < resume_from_scene:
+                continue  # 跳過已處理的 scenes
             seg_wav = os.path.join(temp_dir, f"seg_{idx:04d}.wav")
+            sys.stderr.write(f"[ASR] Scene {idx}: {start_t:.1f}-{end_t:.1f}s\n")
+            sys.stderr.flush()
             # 用 ffmpeg 萃取出該段音訊
+            t0 = time.time()
             cmd = ["ffmpeg", "-y", "-v", "quiet", "-i", video_path,
                    "-ss", str(start_t), "-to", str(end_t),
                    "-ar", "16000", "-ac", "1", seg_wav]
             subprocess.run(cmd, check=False, capture_output=True)
+            sys.stderr.write(f"[ASR] Scene {idx}: ffmpeg took {time.time()-t0:.1f}s\n")
+            sys.stderr.flush()
 
             if not os.path.exists(seg_wav) or os.path.getsize(seg_wav) < 100:
-                continue  # 跳過空音訊
+                sys.stderr.write(f"[ASR] Scene {idx}: empty audio, skipping\n")
+                sys.stderr.flush()
+                continue
 
             try:
+                t1 = time.time()
                 seg_result, seg_info = model.transcribe(
                     seg_wav, beam_size=5,
                     vad_filter=True,
                     vad_parameters=dict(min_silence_duration_ms=500, speech_pad_ms=200),
                 )
-                if transcript_language is None:
-                    transcript_language = seg_info.language
+                sys.stderr.write(f"[ASR] Scene {idx}: transcribe took {time.time()-t1:.1f}s, language={seg_info.language}\n")
+                sys.stderr.flush()
 
                 scene_segments = []
+                seg_language = seg_info.language if seg_info else transcript_language
                 for segment in seg_result:
                     seg_start = start_t + segment.start
                     seg_end = start_t + segment.end
@@ -236,6 +294,7 @@ def run_asr(video_path, output_path, uuid: str = ""):
                         "end": seg_end,
                         "text": segment.text.strip(),
                         "scene_number": scene_idx + 1,
+                        "language": seg_language,
                     })
                     total_segments += 1