app.py

import torch
import gradio as gr
from diffusers import StableAudioPipeline
import scipy.io.wavfile as wavfile
import gc

device = "cuda" if torch.cuda.is_available() else "cpu"

pipe = StableAudioPipeline.from_pretrained(
    "stabilityai/stable-audio-open-1.0",
    torch_dtype=torch.float16 if device == "cuda" else torch.float32
)

pipe = pipe.to(device)
pipe.enable_attention_slicing()


def generate_audio(prompt, duration):
    with torch.no_grad():
        output = pipe(
            prompt=prompt,
            num_inference_steps=100,   # reduced for stability
            audio_end_in_s=duration
        )

    audio = output.audios[0]

    sample_rate = 44100
    file_path = "generated_audio.wav"
    wavfile.write(file_path, sample_rate, audio.T)

    del output
    torch.cuda.empty_cache()
    gc.collect()

    return file_path


iface = gr.Interface(
    fn=generate_audio,
    inputs=[
        gr.Textbox(label="Sound Prompt"),
        gr.Slider(5, 20, value=10, label="Duration (seconds)")
    ],
    outputs=gr.Audio(type="filepath"),
    title="Stable Audio Generator",
)

if __name__ == "__main__":
    iface.launch()