Update app.py

app.py

@@ -10,6 +10,7 @@ import pandas as pd
 from datetime import timedelta
 
 # --- Configuration & Tokens ---
+# Priority: Secret > Hardcoded
 HARDCODED_HF_TOKEN = "PASTE_YOUR_HF_TOKEN_HERE"
 HARDCODED_GEMINI_KEY = ""
 
@@ -127,7 +128,8 @@ if uploaded_file:
                         device = "cuda" if torch.cuda.is_available() else "cpu"
                         st.write(f"🚀 **Loading WhisperX on {device}...**")
                         
-                        # 1. Transcribe
+                        # 1. Transcribe with WhisperX (Faster-Whisper)
+                        batch_size = 16 # Reduce if low VRAM
                         # Use float16 for GPU, int8 for CPU
                         compute_type = "float16" if device == "cuda" else "int8"
                         
@@ -135,14 +137,14 @@ if uploaded_file:
                         
                         st.write("📝 **Transcribing...**")
                         audio = whisperx.load_audio("temp_audio.wav")
-                        result = model.transcribe(audio, batch_size=16)
+                        result = model.transcribe(audio, batch_size=batch_size)
                         
                         # Cleanup VRAM
                         gc.collect()
                         torch.cuda.empty_cache()
                         del model
                         
-                        # 2. Align
+                        # 2. Align (Improves timestamp accuracy)
                         st.write("⏱️ **Aligning Audio...**")
                         model_a, metadata = whisperx.load_align_model(language_code=result["language"], device=device)
                         result = whisperx.align(result["segments"], model_a, metadata, audio, device, return_char_alignments=False)