fix: audio timeout 20s, cap chunks to 3, cap extraction to 30s — fix stuck at 80%

backend/audio_detector.py

@@ -45,9 +45,15 @@ class AudioExtractorAgent:
                 clip.close()
                 return None, 0
 
+            # Cap at 30s — enough for detection, avoids slow extraction on long videos
+            MAX_AUDIO_SEC = 30
+            audio_clip = clip.audio
+            if clip.duration > MAX_AUDIO_SEC:
+                audio_clip = clip.audio.subclipped(0, MAX_AUDIO_SEC)
+
             # Write to temp WAV
             tmp_wav = tempfile.mktemp(suffix=".wav")
-            clip.audio.write_audiofile(
+            audio_clip.write_audiofile(
                 tmp_wav,
                 fps=self.TARGET_SR,
                 nbytes=2,
@@ -277,6 +283,9 @@ class AudioDecisionAgent:
         if not chunks:
             return 0.5
 
+        # Cap at 3 chunks max — Wav2Vec2 is slow on CPU, 30s of audio is enough
+        chunks = chunks[:3]
+
         fake_probs = []
         for chunk in chunks:
             try:

backend/detector.py

@@ -968,7 +968,10 @@ class DeepfakeAuthenticator:
             face_crops_per_frame = face_future.result()
             if audio_future:
                 try:
-                    audio_result = audio_future.result(timeout=30)
+                    # Hard 20s timeout — never block the whole pipeline for audio
+                    audio_result = audio_future.result(timeout=20)
+                except concurrent.futures.TimeoutError:
+                    logger.warning("Audio analysis timed out after 20s — skipping")
                 except Exception as e:
                     logger.warning(f"Audio analysis failed: {e}")