Update app.py

app.py

@@ -65,7 +65,7 @@ DEVICE_ASSERT_PROMPT = None
 DEVICE_ASSERT_LANG = None
 supported_languages = config.languages
 
-def predict(prompt, language, audio_file_pth, mic_file_path, use_mic):
+def predict(prompt, language, audio_file_pth, mic_file_path, use_mic, temperature, repetition_penalty, gpt_cond_len, gpt_cond_chunk_len, max_ref_length):
     try:
         if use_mic:
             speaker_wav = mic_file_path
@@ -75,15 +75,25 @@ def predict(prompt, language, audio_file_pth, mic_file_path, use_mic):
         if len(prompt) < 2 or len(prompt) > 200:
             return None, None, "El texto debe tener entre 2 y 200 caracteres."
 
-        gpt_cond_latent, speaker_embedding = model.get_conditioning_latents(audio_path=speaker_wav, gpt_cond_len=30, gpt_cond_chunk_len=4, max_ref_length=60)
+        # Ajustamos el tamaño del prompt si es necesario
+        max_prompt_length = model.config.max_text_token_len
+        if len(prompt) > max_prompt_length:
+            prompt = prompt[:max_prompt_length]
+
+        gpt_cond_latent, speaker_embedding = model.get_conditioning_latents(
+            audio_path=speaker_wav,
+            gpt_cond_len=gpt_cond_len,
+            gpt_cond_chunk_len=gpt_cond_chunk_len,
+            max_ref_length=max_ref_length
+        )
 
         out = model.inference(
             prompt,
             language,
             gpt_cond_latent,
             speaker_embedding,
-            repetition_penalty=5.0,
-            temperature=0.75,
+            temperature=temperature,
+            repetition_penalty=repetition_penalty,
         )
         
         torchaudio.save("output.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000)
@@ -94,10 +104,10 @@ def predict(prompt, language, audio_file_pth, mic_file_path, use_mic):
         return gr.make_waveform("output.wav"), "output.wav", metrics_text
 
     except Exception as e:
-        print(f"Error: {str(e)}")
+        print(f"Error detallado: {str(e)}")
         return None, None, f"Error: {str(e)}"
 
-# Interfaz de Gradio simplificada
+# Interfaz de Gradio actualizada
 with gr.Blocks(theme=gr.themes.Base()) as demo:
     gr.Markdown("# Sintetizador de Voz XTTS")
     
@@ -111,6 +121,13 @@ with gr.Blocks(theme=gr.themes.Base()) as demo:
             
             use_mic.change(fn=lambda x: gr.update(visible=x), inputs=[use_mic], outputs=[mic_file])
             
+            # Nuevos sliders para parámetros de inferencia
+            temperature = gr.Slider(label="Temperatura", minimum=0.1, maximum=1.0, value=0.75, step=0.01)
+            repetition_penalty = gr.Slider(label="Penalización de repetición", minimum=1.0, maximum=10.0, value=5.0, step=0.1)
+            gpt_cond_len = gr.Slider(label="GPT Cond Len", minimum=10, maximum=50, value=30, step=1)
+            gpt_cond_chunk_len = gr.Slider(label="GPT Cond Chunk Len", minimum=1, maximum=10, value=4, step=1)
+            max_ref_length = gr.Slider(label="Max Ref Length", minimum=30, maximum=120, value=60, step=1)
+            
             generate_button = gr.Button("Generar voz")
         
         with gr.Column():
@@ -120,7 +137,7 @@ with gr.Blocks(theme=gr.themes.Base()) as demo:
     
     generate_button.click(
         predict,
-        inputs=[input_text, language, audio_file, mic_file, use_mic],
+        inputs=[input_text, language, audio_file, mic_file, use_mic, temperature, repetition_penalty, gpt_cond_len, gpt_cond_chunk_len, max_ref_length],
         outputs=[waveform, output_audio, metrics]
     )