Update app.py

app.py

@@ -1,9 +1,4 @@
-# -*- coding: utf-8 -*-
-# POUR GOOGLE COLAB, EXÉCUTEZ CES CELLULES AVANT DE LANCER LE SCRIPT :
-# !apt-get install -y ffmpeg
-# !pip install gradio huggingface_hub torch
-# !pip install git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[all]
-#
+
 
 import os, shlex, subprocess, tempfile, traceback, time, glob, gc, shutil
 import torch
@@ -63,15 +58,22 @@ def get_model(name):
     
     if not nemo_file: raise FileNotFoundError("Fichier .nemo introuvable.")
 
-    from nemo.core.connectors.save_restore_connector import SaveRestoreConnector
-    
-    # Correctif pour les clés "embedding_model" inattendues
-    model = nemo_asr.models.ASRModel.restore_from(
-        nemo_file, 
-        map_location=torch.device(DEVICE),
-        save_restore_connector=SaveRestoreConnector(),
-        strict=False
-    )
+    # Correctif pour les clés "embedding_model" inattendues et erreur __init__
+    try:
+        model = nemo_asr.models.ASRModel.restore_from(
+            nemo_file, 
+            map_location=torch.device(DEVICE),
+            strict=False,
+            override_config_path=None
+        )
+    except Exception as e:
+        print(f"⚠️ Tentative de chargement alternatif : {e}")
+        # Fallback sans override_config_path
+        model = nemo_asr.models.ASRModel.restore_from(
+            nemo_file, 
+            map_location=torch.device(DEVICE),
+            strict=False
+        )
     
     model.to(DEVICE).eval()
     if DEVICE == "cuda":
@@ -166,33 +168,15 @@ def pipeline(video_in, model_name):
         full_wav = os.path.join(tmp_dir, "full.wav")
         subprocess.run(f"ffmpeg -y -threads 0 -i {shlex.quote(video_in)} -vn -ac 1 -ar 16000 {full_wav}", shell=True, check=True)
         
-        yield "⏳ Phase 2/4 : Segmentation Intelligente...", None
+        yield "⏳ Phase 2/4 : Segmentation (5s optimisé Soloni)...", None
+        
+        # Segmentation fixe 5s (optimal pour Soloni V2/V3)
+        subprocess.run(f"ffmpeg -i {full_wav} -f segment -segment_time 5 -c copy {os.path.join(tmp_dir, 'seg_%03d.wav')}", shell=True, check=True)
+        files = sorted(glob.glob(os.path.join(tmp_dir, "seg_*.wav")))
         
-        # Tentative de segmentation intelligente
-        try:
-            cut_points = smart_segment_audio(full_wav, target_duration=5.0)
-        except Exception as e:
-            print(f"Warning smart segment: {e}")
-            cut_points = None
-            
         segment_files = []
-        if cut_points:
-            # Découpage selon les points calculés
-            for i in range(len(cut_points)-1):
-                start = cut_points[i]
-                duration = cut_points[i+1] - start
-                out_name = os.path.join(tmp_dir, f"seg_{i:03d}.wav")
-                subprocess.run(
-                    f"ffmpeg -y -ss {start:.3f} -t {duration:.3f} -i {full_wav} -c copy {out_name}",
-                    shell=True, check=True
-                )
-                segment_files.append({"file": out_name, "start_offset": start})
-        else:
-            # Fallback méthode brute (moins précis mais robuste)
-            subprocess.run(f"ffmpeg -i {full_wav} -f segment -segment_time 5 -c copy {os.path.join(tmp_dir, 'seg_%03d.wav')}", shell=True, check=True)
-            files = sorted(glob.glob(os.path.join(tmp_dir, "seg_*.wav")))
-            for i, f in enumerate(files):
-                segment_files.append({"file": f, "start_offset": i * 5.0})
+        for i, f in enumerate(files):
+            segment_files.append({"file": f, "start_offset": i * 5.0})
 
         yield f"⏳ Phase 3/4 : Chargement de {model_name}...", None
         model = get_model(model_name)