app.py

import spaces
import os
import torchaudio
from audiocraft.models import AudioGen
from audiocraft.data.audio import audio_write
import gradio as gr

OUTPUT_DIR = "audio_files"
DEFAULT_DURATION = 5  # Default duration for audio generation in seconds
os.makedirs(OUTPUT_DIR, exist_ok=True)

model = None  

@spaces.GPU
def generate_audio(descriptions: str, duration: int = DEFAULT_DURATION):    
    global model
    if model is None:
        model = AudioGen.get_pretrained('facebook/audiogen-medium')

    if not os.path.exists('audio_files'):
        os.makedirs('audio_files')
    model.set_generation_params(duration=duration)
    wav = model.generate([descriptions]) 
    results = []
    print(f"Received call: '{descriptions}' duration={duration}")
    for one_wav in wav:
        filename = f'{descriptions}.wav'
        file_path = os.path.join('audio_files', filename)
        audio_write(file_path, one_wav.cpu(), model.sample_rate, strategy="loudness", loudness_compressor=True, add_suffix=False)
        print(f"Generated audio for '{descriptions}'")
        results.append(file_path)
    
    return results[0]



with gr.Blocks() as demo:
    gr.Markdown("# AudioGen Demo")
    with gr.Row():
        descriptions = gr.Textbox(lines=1, label="Enter a description of the audio")
        duration_slider = gr.Slider(minimum=1, maximum=30, value=DEFAULT_DURATION, step=1, label="Duration (seconds)") 
    with gr.Row():
        generate_button = gr.Button("Generate Audio")
    with gr.Row():
        output = gr.Audio(label="Generated Audio")

    generate_button.click(fn=generate_audio, inputs=[descriptions, duration_slider], outputs=output)

demo = demo.queue(max_size=10, status_update_rate="auto")
demo.launch(share=True)