Update app.py

app.py

@@ -1,28 +1,52 @@
-import os
-import re
+import sys
+import io, os, stat
+import subprocess
+import random
+from zipfile import ZipFile
+import uuid
 import time
+import torch
+import torchaudio
+
+# Mantenemos la descarga de MeCab
+os.system('python -m unidic download')
+
+# Mantenemos el acuerdo de CPML
+os.environ["COQUI_TOS_AGREED"] = "1"
+
+import langid
+import base64
+import csv
+from io import StringIO
+import datetime
+import re
+
 import gradio as gr
+from scipy.io.wavfile import write
 from pydub import AudioSegment
+
 from TTS.api import TTS
 from TTS.tts.configs.xtts_config import XttsConfig
 from TTS.tts.models.xtts import Xtts
 from TTS.utils.generic_utils import get_user_data_dir
-from huggingface_hub import hf_hub_download
 
-# Configuración inicial
-os.environ["COQUI_TOS_AGREED"] = "1"
-os.system('python -m unidic download')
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+from huggingface_hub import hf_hub_download
+import os
+from TTS.utils.manage import get_user_data_dir
 
-# Descargar y configurar el modelo
+# Mantenemos la autenticación y descarga del modelo
 repo_id = "Blakus/Pedro_Lab_XTTS"
 local_dir = os.path.join(get_user_data_dir("tts"), "tts_models--multilingual--multi-dataset--xtts_v2")
 os.makedirs(local_dir, exist_ok=True)
 files_to_download = ["config.json", "model.pth", "vocab.json"]
-
 for file_name in files_to_download:
     print(f"Downloading {file_name} from {repo_id}")
+    local_file_path = os.path.join(local_dir, file_name)
     hf_hub_download(repo_id=repo_id, filename=file_name, local_dir=local_dir)
 
+# Cargamos configuración y modelo
 config_path = os.path.join(local_dir, "config.json")
 checkpoint_path = os.path.join(local_dir, "model.pth")
 vocab_path = os.path.join(local_dir, "vocab.json")
@@ -35,17 +59,24 @@ model.load_checkpoint(config, checkpoint_path=checkpoint_path, vocab_path=vocab_
 
 print("Modelo cargado en CPU")
 
-# Funciones auxiliares
-def split_text(text):
-    return re.split(r'(?<=[.!?])\s+', text)
+# Mantenemos variables globales y funciones auxiliares
+DEVICE_ASSERT_DETECTED = 0
+DEVICE_ASSERT_PROMPT = None
+DEVICE_ASSERT_LANG = None
+supported_languages = config.languages
 
-def predict(prompt, language, reference_audio):
+# Función de inferencia usando parámetros predeterminados del archivo de configuración
+def predict(prompt, language, audio_file_pth, mic_file_path, use_mic):
     try:
-        if len(prompt) < 2 or len(prompt) > 600:
-            return None, "El texto debe tener entre 2 y 600 caracteres."
+        if use_mic:
+            speaker_wav = mic_file_path
+        else:
+            speaker_wav = audio_file_pth
 
-        sentences = split_text(prompt)
+        if len(prompt) < 2 or len(prompt) > 200:
+            return None, None, "El texto debe tener entre 2 y 200 caracteres."
 
+        # Usamos los valores de la configuración predeterminada
         temperature = config.inference.get("temperature", 0.75)
         repetition_penalty = config.inference.get("repetition_penalty", 5.0)
         gpt_cond_len = config.inference.get("gpt_cond_len", 30)
@@ -53,101 +84,57 @@ def predict(prompt, language, reference_audio):
         max_ref_length = config.inference.get("max_ref_length", 60)
 
         gpt_cond_latent, speaker_embedding = model.get_conditioning_latents(
-            audio_path=reference_audio,
+            audio_path=speaker_wav,
             gpt_cond_len=gpt_cond_len,
             gpt_cond_chunk_len=gpt_cond_chunk_len,
             max_ref_length=max_ref_length
         )
 
-        start_time = time.time()
-        combined_audio = AudioSegment.empty()
-
-        for sentence in sentences:
-            out = model.inference(
-                sentence,
-                language,
-                gpt_cond_latent,
-                speaker_embedding,
-                temperature=temperature,
-                repetition_penalty=repetition_penalty,
-            )
-            audio_segment = AudioSegment(
-                out["wav"].tobytes(),
-                frame_rate=24000,
-                sample_width=2,
-                channels=1
-            )
-            combined_audio += audio_segment
-            combined_audio += AudioSegment.silent(duration=500)  # 0.5 segundos de silencio
-
-        inference_time = time.time() - start_time
+        out = model.inference(
+            prompt,
+            language,
+            gpt_cond_latent,
+            speaker_embedding,
+            temperature=temperature,
+            repetition_penalty=repetition_penalty,
+        )
         
-        output_path = "output.wav"
-        combined_audio.export(output_path, format="wav")
+        torchaudio.save("output.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000)
 
-        audio_length = len(combined_audio) / 1000  # duración del audio en segundos
-        real_time_factor = inference_time / audio_length
+        metrics_text = f"Tiempo de generación: {out['inference_time']:.2f} segundos\n"
+        metrics_text += f"Factor de tiempo real: {out['inference_time'] / (len(out['wav']) / 24000):.2f}"
 
-        metrics_text = f"Tiempo de generación: {inference_time:.2f} segundos\n"
-        metrics_text += f"Factor de tiempo real: {real_time_factor:.2f}"
-
-        return output_path, metrics_text
+        return gr.make_waveform("output.wav"), "output.wav", metrics_text
 
     except Exception as e:
         print(f"Error detallado: {str(e)}")
-        return None, f"Error: {str(e)}"
-
-# Configuración de la interfaz de Gradio
-supported_languages = ["es", "en"]
-reference_audios = [
-    "serio.wav",
-    "neutral.wav",
-    "alegre.wav",
-]
-
-theme = gr.themes.Soft(
-    primary_hue="blue",
-    secondary_hue="gray",
-).set(
-    body_background_fill='*neutral_100',
-    body_background_fill_dark='*neutral_900',
-)
-
-description = """
-# Sintetizador de voz de Pedro Labattaglia 🎙️
-
-Sintetizador de voz con la voz del locutor argentino Pedro Labattaglia. 
-
-## Cómo usarlo:
-- Elija el idioma (Español o Inglés)
-- Elija un audio de referencia de la lista 
-- Escriba el texto que desea sintetizar
-- Presione generar voz
-"""
-
-# Interfaz de Gradio
-with gr.Blocks(theme=theme) as demo:
-    gr.Markdown(description)
-
-    with gr.Row():
-        gr.Image("https://i1.sndcdn.com/artworks-000237574740-gwz61j-t500x500.jpg", label="", show_label=False, width=250, height=250)
+        return None, None, f"Error: {str(e)}"
 
+# Interfaz de Gradio actualizada sin sliders
+with gr.Blocks(theme=gr.themes.Base()) as demo:
+    gr.Markdown("# Sintetizador de Voz XTTS")
+    
     with gr.Row():
-        with gr.Column(scale=2):
-            language_selector = gr.Dropdown(label="Idioma", choices=supported_languages)
-            reference_audio = gr.Dropdown(label="Audio de referencia", choices=reference_audios)
+        with gr.Column():
             input_text = gr.Textbox(label="Texto a sintetizar", placeholder="Escribe aquí el texto que quieres convertir a voz...")
-            generate_button = gr.Button("Generar voz", variant="primary")
-
-        with gr.Column(scale=1):
-            generated_audio = gr.Audio(label="Audio generado", interactive=False)
-            metrics_output = gr.Textbox(label="Métricas", value="Tiempo de generación: -- segundos\nFactor de tiempo real: --")
-
+            language = gr.Dropdown(label="Idioma", choices=supported_languages, value="es")
+            audio_file = gr.Audio(label="Audio de referencia", type="filepath")
+            use_mic = gr.Checkbox(label="Usar micrófono")
+            mic_file = gr.Audio(label="Grabar con micrófono", source="microphone", type="filepath", visible=False)
+            
+            use_mic.change(fn=lambda x: gr.update(visible=x), inputs=[use_mic], outputs=[mic_file])
+            
+            generate_button = gr.Button("Generar voz")
+        
+        with gr.Column():
+            output_audio = gr.Audio(label="Audio generado")
+            waveform = gr.Image(label="Forma de onda")
+            metrics = gr.Textbox(label="Métricas")
+    
     generate_button.click(
         predict,
-        inputs=[input_text, language_selector, reference_audio],
-        outputs=[generated_audio, metrics_output]
+        inputs=[input_text, language, audio_file, mic_file, use_mic],
+        outputs=[waveform, output_audio, metrics]
     )
 
-if __name__ == "__main__":
-    demo.launch()
+demo.launch(debug=True)