Fix: handle Whisper transcribe failure for instrumental-only audio (fallback to empty transcript)

scripts/analyze.py

@@ -47,19 +47,31 @@ def extract_audio(video_path):
 
 
 def transcribe_full(audio_path, lang=None, model_size="small"):
-    """Whisper transcript celega avdia. lang=None → auto-detect."""
+    """Whisper transcript celega avdia. lang=None → auto-detect.
+    
+    Vrne empty transcript če Whisper ne najde govora (popolnoma instrumental)."""
     from faster_whisper import WhisperModel
 
     print(f"🧠 Whisper {model_size}, lang={lang or 'auto'}", file=sys.stderr)
     m = WhisperModel(model_size, device="cpu", compute_type="int8")
-    segs, info = m.transcribe(
-        audio_path,
-        language=lang,
-        word_timestamps=True,
-        vad_filter=True,
-    )
-    detected_lang = info.language
-    detected_prob = float(info.language_probability)
+    try:
+        segs, info = m.transcribe(
+            audio_path,
+            language=lang,
+            word_timestamps=True,
+            vad_filter=True,
+        )
+        detected_lang = info.language
+        detected_prob = float(info.language_probability)
+    except (ValueError, RuntimeError) as e:
+        # Whisper failure (např. pri popolnoma instrumentalnih datotekah z VAD)
+        print(f"   ⚠️ Whisper transcribe failed: {e}", file=sys.stderr)
+        return {
+            "language": "unknown",
+            "language_probability": 0.0,
+            "segments": [],
+        }
+
     print(f"   Detekcija: {detected_lang} (p={detected_prob:.2f})", file=sys.stderr)
 
     segments = []