Update app.py

app.py

@@ -115,41 +115,41 @@ async def stream_generator(file_path: str, display_name: str, api_key: str, mode
 
         yield log_msg(f"Ready. Starting generation for full duration...")
 
-        # 4. Prompt Construction (Reverted to Stable Version)
+        # 4. Prompt Construction (Optimized for Sync)
         end_time_instruction = total_duration_str if total_duration_str != "Unknown" else "the absolute end of the file"
         
         instruction_set = f"""
-          1. Transcribe audio from 00:00:00 to {end_time_instruction}.
+          1. Transcribe the audio from 00:00:00 to {end_time_instruction}.
           2. Translate to natural Burmese (Myanmar).
-          3. Output format: Standard SRT.
+          3. **TIMING RULE**:
+             - **Do NOT fill gaps.** If there is music without lyrics, leave it blank.
+             - Timestamps must match the **exact start and end** of the singing/speech.
+             - Do NOT make timestamps continuous (e.g., if line 1 ends at 00:05 and line 2 starts at 00:10, do NOT extend line 1 to 00:10).
+             - Break subtitles when the singer takes a breath.
         """
 
         if include_songs:
-            instruction_set += "\n          4. For SONGS: Use 🎶."
+            instruction_set += "\n          4. SONGS: Capture lyrics. Use 🎶 at start/end."
         else:
             instruction_set += "\n          4. Ignore background music/lyrics."
 
         if include_on_screen_text:
-            instruction_set += "\n          5. For ON-SCREEN TEXT: Use 📝."
+            instruction_set += "\n          5. ON-SCREEN TEXT: Use 📝."
         else:
             instruction_set += "\n          5. Ignore visual text."
 
-        # Stability Instructions
         instruction_set += f"""
-          6. **SYNC CRITICAL**:
-             - The file duration is EXACTLY {end_time_instruction}.
-             - Do NOT stop early. If there is silence, wait for the next dialogue.
-             - Only write "End" when the timestamp reaches {end_time_instruction}.
-
-          7. **TIMESTAMP CORRECTION**:
-             - Start time must be 00:00:00.
-             - **DO NOT** write "01:00:00" (1 hour) unless the video is actually longer than 60 minutes.
-             - e.g., Use "00:01:10", NOT "01:01:10".
+          6. **SYNC CHECK**:
+             - Start time MUST be 00:00:00.
+             - Ensure timestamps do NOT overlap.
+             - Continue until {end_time_instruction}.
+             - **Timestamp Format**: HH:MM:SS,mmm (e.g. 00:01:23,450)
+             - **NO HALLUCINATED HOURS**: Do not write "01:00:00" if video is short.
         """
 
         prompt = f"""
-          Task: Create Burmese SRT subtitles for the ENTIRE duration.
-          TOTAL DURATION: {total_duration_str}
+          Task: Create accurate, synced Burmese SRT subtitles.
+          Total Duration: {total_duration_str}
           INSTRUCTIONS:{instruction_set}
         """
 
@@ -161,7 +161,7 @@ async def stream_generator(file_path: str, display_name: str, api_key: str, mode
         response_stream = model.generate_content(
             [video_file, prompt],
             stream=True,
-            generation_config={"temperature": 0.2} # Reverted to 0.2 (Stable)
+            generation_config={"temperature": 0.0} # Temperature 0.0 for strictest adherence to timing
         )
 
         chunk_count = 0