Update app.py

app.py

@@ -6,6 +6,7 @@ import os
 import requests
 import torch
 import numpy
+import torchaudio
 from datetime import timedelta
 from pyannote.audio import Pipeline
 from huggingface_hub import hf_hub_download
@@ -164,13 +165,18 @@ if uploaded_file:
                         
                         pipeline = Pipeline.from_pretrained(config_path)
                         
-                        # Use GPU if available, but fallback to CPU if issues persist
+                        # Load audio manually into tensor for robust processing
+                        waveform, sample_rate = torchaudio.load("temp_audio.wav")
+                        
+                        # Ensure we are on GPU
                         if torch.cuda.is_available():
                             st.write("🚀 Using GPU for Diarization")
                             pipeline.to(torch.device("cuda"))
+                            waveform = waveform.to(torch.device("cuda"))
                         
-                        # Run on WAV file
-                        diarization_output = pipeline("temp_audio.wav")
+                        # Run pipeline by passing the tensor dictionary
+                        # This is the most reliable way to avoid file I/O issues in Docker
+                        diarization_output = pipeline({"waveform": waveform, "sample_rate": sample_rate})
                         
                         # Handle Wrapper
                         if isinstance(diarization_output, tuple):
@@ -178,7 +184,7 @@ if uploaded_file:
                         else:
                             diarization = diarization_output
                             
-                        # Extract Annotation
+                        # CRITICAL FIX: Extract annotation from wrapper
                         if not hasattr(diarization, "itertracks"):
                              if hasattr(diarization_output, "annotation"):
                                  diarization = diarization_output.annotation
@@ -221,6 +227,7 @@ if uploaded_file:
                         mid_time = (segment['start'] + segment['end']) / 2
                         speaker = "Unknown"
                         
+                        # Matching logic
                         if speaker_turns:
                             # 1. Strict
                             for turn in speaker_turns:
@@ -263,6 +270,7 @@ if uploaded_file:
                 st.error("Gemini API Key required. Please set it in Secrets or app.py.")
             else:
                 with st.spinner("Analyzing..."):
+                    # Fixed missing parenthesis here
                     edl_segments = call_gemini_for_edl(st.session_state.transcript, brief, ACTIVE_GEMINI_KEY)
                     if edl_segments:
                         final_edl = generate_cmx_edl("AI_Senior_Editor_Cut", edl_segments, uploaded_file.name, fps)