Create app.py

app.py

@@ -20,12 +20,16 @@ def text_to_speech(text):
     mel_output, mel_length, alignment = tacotron2.encode_text(text)
     
     # Decode mel spectrogram to waveform
-    # FIX: hifi_gan.decode_batch() returns only 1 value
     waveforms = hifi_gan.decode_batch(mel_output)
     
+    # Convert waveform to numpy and normalize to int16 range
+    waveform = waveforms.squeeze(1).cpu().numpy()
+    waveform = waveform / max(abs(waveform))  # Normalize to range [-1, 1]
+    waveform = (waveform * 32767).astype("int16")  # Scale to int16 range
+    
     # Save waveform as audio file
     audio_path = "output.wav"
-    write(audio_path, 22050, waveforms.squeeze(1).cpu().numpy())
+    write(audio_path, 22050, waveform)
     return audio_path
 
 # Streamlit UI