Update app.py

app.py

@@ -54,7 +54,9 @@ def predict(text, speaker):
     speech = model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder)
     #speech = model.generate_speech(inputs["input_ids"], speaker_embedding, vocoder=vocoder)
 
-    speech = (speech.numpy() * 32767).astype(np.int16)
+    #speech = (speech.numpy() * 32767).astype(np.int16)
+    speech = np.tanh(speech) * 32767  # Non-linear scaling to reduce harshness
+    speech = speech.astype(np.int16)
     return (16000, speech)