fix: return actual audio data (numpy) from ZeroGPU function instead of file paths to ensure Gradio receives the content before the worker environment is destroyed

app.py

@@ -104,7 +104,14 @@ def _full_pipeline_gpu(song_file, reference_path, pitch, diffusion_steps, simila
         instrumental_volume=float(instrumental_volume),
     )
 
-    return vocals_path, converted_path, final_path
+    import librosa
+    # Load back the audio data to return it directly.
+    # This bypasses ZeroGPU filesystem sync issues.
+    v_data, v_sr = librosa.load(vocals_path, sr=None)
+    c_data, c_sr = librosa.load(converted_path, sr=None)
+    f_data, f_sr = librosa.load(final_path, sr=None)
+
+    return (v_sr, v_data), (c_sr, c_data), (f_sr, f_data)
 
 
 def train_voice_model(audio_file, model_name, progress=gr.Progress()):
@@ -357,18 +364,15 @@ with gr.Blocks(
                     preview_vocals = gr.Audio(
                         label="Voz original (separada)",
                         interactive=False,
-                        type="filepath",
                     )
                     preview_converted = gr.Audio(
                         label="Voz convertida",
                         interactive=False,
-                        type="filepath",
                     )
                     gr.Markdown("**Resultado final:**")
                     final_output = gr.Audio(
                         label="Canción final (voz + instrumentos)",
                         interactive=False,
-                        type="filepath",
                     )
 
             refresh_btn.click(