app.py

import gradio as gr
from audio_style_transfer.models import timedomain

def audioStyleTransfer(content,style):
    print (style,content)
    output = "/tmp/outputfile.wav"
    sr = librosa.get_samplerate(content)
    timedomain.run( style_fname=style,
        content_fname=content,
        output_fname=output,
        norm=False,
        n_fft=8192,          # 512 to sr / 2. Higher is better quality but is slower.
        n_layers=2,          # 1 to 3. Higher is better quality but is slower.
        n_filters=4096,      # 512 - 4096. Higher is better quality but is slower.
        hop_length=256,      # 256 to n_fft / 2. The lower this value, the better the temporal resolution.
        alpha=0.0005,        # 0.0001 to 0.01. The higher this value, the more of the original "content" bleeds                   
        k_w=4,               # 3 to 5. The higher this value, the more complex the patterns it can synthesize.
        iterations=300,      # 100 to 1000. Higher is better quality but is slower.
        stride=1,            # 1 to 3. Lower is better quality but is slower.
        sr=sr)
    print ("output is " ,output)
    return  output

iface = gr.Interface(fn=audioStyleTransfer, title="Time Domain Audio Style transfer",
                    description="Forked from https://github.com/pkmital/time-domain-neural-audio-style-transfer Built to style transfer to audio using style audio.\
       it seems to work best for shorter clips", inputs=[gr.Audio(source="upload",type="filepath",label="Content"),gr.Audio(source="upload",type="filepath",label="Style")], outputs=gr.Audio(label="Output"))
iface.launch()