| import torch | |
| 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() |