float coinversion therpy

app.py

@@ -104,9 +104,11 @@ def text_to_speech(text):
     # Make sure speech is float32 before any conversion to avoid the error
     audio_array = speech.cpu().numpy().astype(np.float32)
     
+    # Normalize and convert to int16 for output
+    audio_array = (audio_array / np.max(np.abs(audio_array) + 1e-6) * 32767).astype(np.int16)
     
     # Reshape for fastrtc
-    audio_array = audio_array.reshape(1, -1)
+    audio_array = audio_array.reshape(1, -1).astype(np.float32)
     
     return (16000, audio_array)  # SpeechT5 uses 16kHz sample rate