Update app.py

app.py

@@ -1,42 +1,57 @@
-import gradio as gr
-import random
 import torch
-from audiocraft.models import MusicGen
-from audiocraft.data.audio import audio_write
-import spaces
-
-model = MusicGen.get_pretrained("small")
-
-@spaces.GPU()
-def infer(duration, descriptions):
-    if len(descriptions) > 8192:
-        return "Error: La descripción no puede exceder los 8192 caracteres."
-
-    seed = random.randint(0, 10000)
-    torch.manual_seed(seed)
-
-    model.set_generation_params(duration=duration, temperature=0.5)
-    wav = model.generate(descriptions.split(", "))
-    output_files = []
-
-    for idx, one_wav in enumerate(wav):
-        file_name = f'output_{idx}.wav'
-        audio_write(file_name, one_wav.cpu(), model.sample_rate, strategy="loudness")
-        output_files.append(file_name)
-
-    return output_files[0]
-
-with gr.Blocks() as demo:
-    gr.Markdown("# Generador de Música con MusicGen")
-    duration_input = gr.Slider(minimum=1, maximum=600, label="Duración de la canción (segundos)", value=8)
-    description_input = gr.Textbox(placeholder="Ejemplo: happy rock, energetic EDM", label="Descripción de la música")
-    generate_button = gr.Button("Generar Música")
-    output_audio = gr.Audio(label="Escuchar Música", type="filepath")
-
-    generate_button.click(
-        fn=infer,
-        inputs=[duration_input, description_input],
-        outputs=output_audio
-    )
-
-demo.launch()
+import gradio as gr
+from transformers import AutoProcessor, MusicgenForConditionalGeneration
+from TTS.api import TTS
+import librosa
+import numpy as np
+
+processor = AutoProcessor.from_pretrained("facebook/musicgen-small")
+model_musicgen = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")
+model_tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cpu")
+
+def generate_music(text_prompt, audio_prompt=None, tts_text=None):
+    inputs = processor(text=text_prompt, padding=True, return_tensors="pt")
+    max_duration_sec = 600
+    min_duration_sec = 30
+    max_new_tokens = int(max_duration_sec * model_musicgen.config.audio_length // model_musicgen.config.generation_length)
+    min_new_tokens = int(min_duration_sec * model_musicgen.config.audio_length // model_musicgen.config.generation_length)
+    audio_values = model_musicgen.generate(**inputs.to("cpu"), max_new_tokens=max_new_tokens, min_new_tokens=min_new_tokens)
+    music = audio_values.cpu().numpy()[0]
+
+    if audio_prompt is not None and tts_text is not None:
+        cloned_audio = model_tts.tts_with_vc(tts_text, speaker_wav=audio_prompt, language="es")
+
+        if model_tts.synthesizer.output_sample_rate != 44100:
+            cloned_audio, _ = librosa.resample(cloned_audio, model_tts.synthesizer.output_sample_rate, 44100)
+
+        try:
+            music_np = librosa.util.buf_to_float(music, n_bytes=2)
+
+            if len(music_np) > len(cloned_audio):
+                padding = np.zeros(len(music_np) - len(cloned_audio))
+                cloned_audio = np.concatenate([cloned_audio, padding])
+            elif len(cloned_audio) > len(music_np):
+                padding = np.zeros(len(cloned_audio) - len(music_np))
+                music_np = np.concatenate([music_np, padding])
+
+            combined_audio = music_np + cloned_audio
+            return (44100, combined_audio)
+        except ImportError:
+            print("Error: Se requiere librosa para combinar audio. Instale con 'pip install librosa'")
+            return (44100, music)
+
+    return (44100, music)
+
+iface = gr.Interface(
+    fn=generate_music,
+    inputs=[
+        gr.Textbox(label="Descripción de la música"),
+        gr.Audio(source="microphone", type="filepath", label="Audio de voz (Opcional)", optional=True),
+        gr.Textbox(label="Texto para clonar la voz (Opcional)", optional=True),
+    ],
+    outputs=gr.Audio(label="Música generada", type="numpy"),
+    title="Generador de Música con MusicGen y XTTS",
+    description="Introduce una descripción de la música que deseas generar y opcionalmente un audio de voz para clonar con texto.",
+)
+
+iface.launch()