Update app.py

app.py

@@ -1,5 +1,6 @@
 import gradio as gr
 from bark import SAMPLE_RATE, generate_audio, preload_models
+from bark.generation import generate_text_semantic, text_to_semantic
 from scipy.io.wavfile import write as write_wav
 import tempfile
 import librosa
@@ -39,20 +40,28 @@ def preprocess_audio_to_npz(audio_path):
     # Ensure audio is a float32 array
     audio = audio.astype(np.float32)
     
-    # Generate semantic tokens directly using Bark's internal processing
-    # Since HuBERT models are not implemented, we rely on generate_audio's history prompt
-    # This is a simplified approach assuming Bark can handle raw audio for history prompt
     with torch.device("cpu"):
-        # Generate audio tokens to create a history prompt
-        # We use a dummy text to generate a history prompt from the audio
+        # Generate semantic tokens from the audio
+        # Use a dummy text to initialize the semantic token generation
         dummy_text = "Dummy text for history prompt generation."
-        audio_array = generate_audio(dummy_text, history_prompt=audio_path)
+        semantic_tokens = text_to_semantic(dummy_text, temp=0.7, silent=True)
         
-        # Save the audio array as a temporary .npz file
+        # Generate coarse tokens from semantic tokens
+        coarse_tokens = generate_text_semantic(
+            semantic_tokens=semantic_tokens,
+            max_gen_len=512,
+            temp=0.7,
+            silent=True
+        )
+        
+        # Create history prompt dictionary
         history_prompt = {
-            "audio": audio_array
+            "semantic_prompt": semantic_tokens,
+            "coarse_prompt": coarse_tokens,
+            "fine_prompt": coarse_tokens  # Fine prompt is often same as coarse for Bark
         }
         
+        # Save to temporary .npz file
         with tempfile.NamedTemporaryFile(suffix=".npz", delete=False) as temp_file:
             np.savez(temp_file.name, **history_prompt)
             npz_path = temp_file.name