app.py

from spleeter.separator import Separator
import gradio as gr
import shutil
import librosa
import soundfile
import os
from gradio_client import Client, file
import subprocess

from constant import weights, indices

try:
  shutil.rmtree("output")
except FileNotFoundError:
  pass


def spleeter(separator, aud, model_dropdown, model_file, index_file, pitch):
  filename = os.path.basename(aud).split('.')[0]
  accompaniment_filename = f"./output/audio_example/{filename}_accompaniment.wav"
  vocal_filename = f"./output/audio_example/{filename}_vocals.wav"
  song_filename = f'./output/audio_example/{filename}_song.wav'

  separator.separate_to_file(aud, "output/", filename_format="audio_example/%s_{instrument}.wav" % filename)

  # change pitch
  y, sr = librosa.load(accompaniment_filename)
  new_accompaniment = librosa.effects.pitch_shift(y, sr, pitch)
  soundfile.write(accompaniment_filename, new_accompaniment, sr, )

  # send vocal for processing
  # https://huggingface.co/spaces/r3gm/rvc_zero

  model = model_file if model_dropdown is None else weights[model_dropdown]
  index = index_file if model_dropdown is None else indices[model_dropdown]

  client = Client("r3gm/rvc_zero")
  result = client.predict(
    audio_files=[file(vocal_filename)],
    file_m=file(model),
    pitch_alg="rmvpe+",
    pitch_lvl=pitch,
    file_index=file(index),
    index_inf=0.75,
    r_m_f=3,
    e_r=0.25,
    c_b_p=0.5,
    active_noise_reduce=False,
    audio_effects=False,
    api_name="/run"
  )[0]

  # combine accompaniment and vocal
  command = f"""ffmpeg -i "{result}" -i "{accompaniment_filename}" -filter_complex amix=inputs=2:duration=longest "{song_filename}" """
  subprocess.call(command, shell=True)

  return song_filename

inputs = [
          gr.Audio(sources='upload', label="Input Audio File", type="filepath"),
    gr.Dropdown(choices=list(weights.keys()), value=list(weights.keys()), label="existing people model"),
gr.File(
        label="Model file (use this if your existing people model does not exist in the dropdown)",
        type="filepath",
        height=130,
    ),
gr.File(
        label="Index file (use this if your existing people model does not exist in the dropdown)",
        type="filepath",
        height=130,
    ),
gr.Slider(
        label="Pitch level",
        minimum=-24,
        maximum=24,
        step=1,
        value=0,
        visible=True,
        interactive=True,
    )
]
outputs =  [
  gr.Audio(label="Output Audio", type="filepath"),
]

title = "Music Spleeter"
description = "Clearing a musical composition of the performer's voice is a common task. It is solved well, for example, by professional audio file editing programs. AI algorithms have also been gaining ground recently."
article = "<div style='text-align: center; max-width:800px; margin:10px auto;'><p>In this case we use Deezer's Spleeter with ready pretrained models. It can leave as an output both just the music and just the performer's voice.</p><p>Sources: <a href='https://github.com/deezer/spleeter/' target='_blank'>Spleeter</a>: a Fast and Efficient Music Source Separation Tool with Pre-Trained Models</p><p style='text-align: center'><a href='https://starstat.yt/cat/music' target='_blank'>StarStat Music</a>: Youtubers Net Worth in category Music</p></div>"


if __name__ == '__main__':
  separator = Separator('spleeter:2stems')
  lambda_spleeter = lambda aud, model_dropdown, model_file, index_file, pitch: spleeter(separator, aud=aud, model_dropdown=model_dropdown, model_file=model_file, index_file=index_file, pitch=pitch)
  gr.Interface(lambda_spleeter, inputs, outputs).launch()