Update app.py

app.py

@@ -16,16 +16,21 @@ model = BetterTransformer.transform(model, keep_original_model=False)
 # Enable CPU offload
 model.enable_cpu_offload()
 
-from transformers import AutoProcessor
+import numpy as np
+from scipy.io import wavfile
 
 def infer(text_prompt):
     text_prompt = text_prompt
     inputs = processor(text_prompt).to(device)
     with torch.inference_mode():
         speech_output = model.generate(**inputs, do_sample = True, fine_temperature = 0.4, coarse_temperature = 0.8)
-        #audio_out = speech_output[0].cpu().numpy()
-        audio_out = speech_output[0]
-    return audio_out
+        audio_out = speech_output[0].cpu().numpy()
+
+    # Assuming audio_out contains audio data and the sampling rate
+    sampling_rate = 44100  # Replace with your actual sampling rate
+    wavfile.write("output.wav", sampling_rate, audio_out)
+    
+    return "output.wav"
 
 with gr.Blocks() as demo:
     with gr.Column():