app.py

import torch
import os
import soundfile as sf
from diffusers import StableAudioPipeline
from huggingface_hub import login

# Retrieve the token securely from environment variables
HUGGINGFACE_TOKEN = os.getenv("HF_TOKEN")
if HUGGINGFACE_TOKEN is None:
    raise ValueError("Missing Hugging Face API token. Set 'HF_TOKEN' in Secrets.")

# Authenticate with Hugging Face Hub
login(HUGGINGFACE_TOKEN)

# Set up the device
device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if device == "cuda" else torch.float32

# Load the model
pipe = StableAudioPipeline.from_pretrained(
    "stabilityai/stable-audio-open-1.0",
    torch_dtype=torch_dtype,
    use_auth_token=True  # Token provided via 'login' earlier
)
pipe = pipe.to(device)

# Audio generation function
def generate_audio(prompt, negative_prompt, duration, seed):
    generator = torch.Generator(device).manual_seed(seed)
    audio_output = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=50,
        audio_end_in_s=duration,
        num_waveforms_per_prompt=1,
        generator=generator
    ).audios
    
    # Save the generated audio
    output_audio = audio_output[0].T.float().cpu().numpy()
    sf.write("output.wav", output_audio, pipe.vae.sampling_rate)
    return "output.wav"

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("## 🎧 Stable Audio Open - Interactive Audio Generation 🎼")
    with gr.Row():
        prompt_input = gr.Textbox(label="Prompt", value="The sound of a hammer hitting a wooden surface.")
        negative_input = gr.Textbox(label="Negative Prompt", value="Low quality.")
    with gr.Row():
        duration_input = gr.Slider(minimum=1, maximum=10, step=0.5, value=2, label="Duration (seconds)")
        seed_input = gr.Number(label="Random Seed", value=0)
    generate_button = gr.Button("Generate Audio")
    output_audio = gr.Audio(label="Generated Audio", type="filepath")
    
    generate_button.click(
        generate_audio, 
        inputs=[prompt_input, negative_input, duration_input, seed_input], 
        outputs=output_audio
    )

demo.launch()