Update app.py

app.py

@@ -1,7 +1,15 @@
+import os
+import numpy as np
+
+# --- NUMPY 2.0 PATCH ---
+# This must run before any other library imports to prevent crashes
+# with pyannote/whisperx which might expect the old 'np.NaN' attribute.
+if not hasattr(np, 'NaN'):
+    np.NaN = np.nan
+
 import streamlit as st
 import subprocess
 import json
-import os
 import requests
 import torch
 import whisperx
@@ -92,7 +100,8 @@ with st.sidebar:
     fps = st.number_input("Timeline FPS", value=25)
     
     st.header("Model Settings")
-    model_size = st.selectbox("Whisper Model", ["large-v2", "medium", "base"], index=0)
+    # T4 has 16GB VRAM, large-v2 works well
+    model_size = st.selectbox("Whisper Model", ["large-v2", "medium"], index=0)
     num_speakers = st.number_input("Speakers (0=Auto)", min_value=0, value=0)
     
     st.divider()
@@ -111,7 +120,10 @@ if uploaded_file:
             if not ACTIVE_HF_TOKEN or "PASTE_YOUR_HF_TOKEN" in ACTIVE_HF_TOKEN:
                 st.error("Please provide a valid Hugging Face Token.")
             else:
-                with st.spinner("Processing... This may take a moment."):
+                status_container = st.empty()
+                with status_container.container():
+                    st.write("🔄 **Processing Started...**")
+                    
                     # Save local temp file
                     with open("temp_input", "wb") as f:
                         f.write(uploaded_file.getbuffer())
@@ -126,10 +138,12 @@ if uploaded_file:
                     
                     try:
                         device = "cuda" if torch.cuda.is_available() else "cpu"
+                        if device == "cpu":
+                            st.warning("⚠️ No GPU detected. WhisperX will be very slow.")
+                            
                         st.write(f"🚀 **Loading WhisperX on {device}...**")
                         
-                        # 1. Transcribe with WhisperX (Faster-Whisper)
-                        batch_size = 16 # Reduce if low VRAM
+                        # 1. Transcribe
                         # Use float16 for GPU, int8 for CPU
                         compute_type = "float16" if device == "cuda" else "int8"
                         
@@ -137,14 +151,14 @@ if uploaded_file:
                         
                         st.write("📝 **Transcribing...**")
                         audio = whisperx.load_audio("temp_audio.wav")
-                        result = model.transcribe(audio, batch_size=batch_size)
+                        result = model.transcribe(audio, batch_size=16)
                         
                         # Cleanup VRAM
                         gc.collect()
                         torch.cuda.empty_cache()
                         del model
                         
-                        # 2. Align (Improves timestamp accuracy)
+                        # 2. Align (Improves timestamp accuracy for diarization)
                         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)
@@ -170,10 +184,11 @@ if uploaded_file:
                         
                         # Format for Gemini
                         processed_segments = []
+                        # WhisperX structure is slightly different, it returns 'segments' list
                         for segment in final_result["segments"]:
                             processed_segments.append({
                                 "speaker": segment.get("speaker", "Unknown"),
-                                "text": segment["text"],
+                                "text": segment["text"].strip(),
                                 "start": segment["start"],
                                 "end": segment["end"]
                             })