Update app.py

app.py

@@ -96,7 +96,13 @@ def synthesize_speech(text, ref_audio, ref_text):
         sf.write(temp_audio.name, audio_data, samplerate=sample_rate, format='WAV')
         temp_audio.flush()
     
-    # Use the **original** signature – never change it
+    # Load with soundfile (not torchaudio) and convert to tensor
+    wav, sr = sf.read(temp_audio.name)
+    wav = torch.from_numpy(wav).float()
+    if sr != 24000:
+        wav = torchaudio.functional.resample(wav, sr, 24000)
+    
+    # Feed tensor directly to model – never touches torchaudio.load
     out = model(text, ref_audio_path=temp_audio.name, ref_text=ref_text)
              
     # Normalize output and save