app.py

from diffusers import AudioLDMPipeline
import torch
import scipy
import gradio as gr

repo_id = "cvssp/audioldm-m-full"
pipe = AudioLDMPipeline.from_pretrained(repo_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

def my_inference_function(prompt):
	user_prompt = prompt
	#prompt = "a woman singing, dry, sample, one shot, a minor, mastered, legato, clear, dry, high quaility"
	neg_prompt = "low quality, average quality, noisy, distortion, realistic"
	seed = 51
	steps = 5
	guidance = 0.1
	length = 10

	g_cuda = torch.Generator(device='cuda')
	g_cuda.manual_seed(seed)
	audio = pipe(user_prompt,
             num_inference_steps=steps,
             audio_length_in_s=length,
             negative_prompt=neg_prompt,
             guidance_scale=guidance,
             num_waveforms_per_prompt=1,
             generator = g_cuda
            ).audios[0]
	return gr.make_waveform((16000,audio))
	#return scipy.io.wavfile.write(user_prompt+"_s"+str(steps)+"_g"+str(guidance)+"_"+str(length)+"sec_"+str(seed)+".wav", rate=16000, data=audio)

gradio_interface = gr.Interface(
	fn = my_inference_function,
	inputs = "text",
	outputs = "audio",
)

gradio_interface.launch()