Update app.py

app.py

@@ -5,12 +5,32 @@ import json
 import os
 import requests
 import torch
-import numpy
+import numpy as np
 from datetime import timedelta
 from pyannote.audio import Pipeline
 from huggingface_hub import login, hf_hub_download
 from pydub import AudioSegment
 
+# --- Safe Globals for PyTorch 2.6+ ---
+try:
+    from pyannote.audio.core.task import Specifications, Problem, Resolution
+    from pyannote.audio.core.model import Model
+    from pyannote.audio.pipelines.speaker_diarization import SpeakerDiarization
+    
+    torch.serialization.add_safe_globals([
+        torch.torch_version.TorchVersion,
+        Specifications,
+        Problem,
+        Resolution,
+        Model,
+        SpeakerDiarization,
+        np.dtype,
+        torch.nn.modules.container.ModuleList,
+        np.core.multiarray.scalar
+    ])
+except Exception as e:
+    print(f"Safe Globals Warning: {e}")
+
 # --- Configuration & Tokens ---
 HARDCODED_HF_TOKEN = "PASTE_YOUR_HF_TOKEN_HERE"
 HARDCODED_GEMINI_KEY = ""
@@ -127,13 +147,12 @@ if uploaded_file:
                     st.write("🗣️ **Running Speaker Diarization...**")
                     diarization = None
                     try:
-                        # SIMPLEST LOADING METHOD (v2.1.1 compatible)
-                        # We use hf_hub_download to get the config.yaml path manually
-                        # This avoids "use_auth_token" ambiguity in From_Pretrained
+                        # FIX: Pass revision separately, remove @ from repo_id
                         config_path = hf_hub_download(
-                            repo_id="pyannote/speaker-diarization@2.1", 
+                            repo_id="pyannote/speaker-diarization", 
+                            revision="2.1",  # Explicit revision
                             filename="config.yaml", 
-                            token=ACTIVE_HF_TOKEN # FIXED ARGUMENT NAME
+                            token=ACTIVE_HF_TOKEN
                         )
                         
                         pipeline = Pipeline.from_pretrained(config_path)
@@ -162,31 +181,26 @@ if uploaded_file:
                     
                     if diarization:
                         # 2.1.1 returns a proper Annotation object directly
-                        try:
-                            for turn, _, speaker_id in diarization.itertracks(yield_label=True):
-                                speaker_turns.append({"start": turn.start, "end": turn.end, "speaker": speaker_id})
-                            
-                            if len(speaker_turns) > 0:
-                                st.write(f"✅ Found {len(speaker_turns)} speaker turns.")
-                            else:
-                                st.warning("⚠️ Pipeline ran but returned no tracks.")
-                        except AttributeError:
-                             st.error("Could not iterate tracks. Output object format mismatch.")
+                        for turn, _, speaker_id in diarization.itertracks(yield_label=True):
+                            speaker_turns.append({"start": turn.start, "end": turn.end, "speaker": speaker_id})
+                        
+                        if len(speaker_turns) > 0:
+                            st.write(f"✅ Found {len(speaker_turns)} speaker turns.")
+                        else:
+                            st.warning("⚠️ Pipeline ran but returned no tracks.")
 
                     for segment in result['segments']:
                         mid_time = (segment['start'] + segment['end']) / 2
                         speaker = "Unknown"
                         
                         if speaker_turns:
-                            # 1. Strict overlap
                             for turn in speaker_turns:
                                 if turn["start"] <= mid_time <= turn["end"]:
                                     speaker = turn["speaker"]
                                     break
                             
-                            # 2. Relaxed distance fallback (1 sec)
                             if speaker == "Unknown":
-                                best_dist = 1.0
+                                best_dist = 1.0 
                                 for turn in speaker_turns:
                                     dist = min(abs(turn["start"] - mid_time), abs(turn["end"] - mid_time))
                                     if dist < best_dist: