app.py

# !git clone https://github.com/Edresson/Coqui-TTS -b multilingual-torchaudio-SE TTS

import os
import shutil
import gradio as gr

import sys

import string
import time
import argparse
import json

import numpy as np
# import IPython
# from IPython.display import Audio

import torch

from TTS.tts.utils.synthesis import synthesis
from TTS.tts.utils.text.symbols import make_symbols, phonemes, symbols
try:
  from TTS.utils.audio import AudioProcessor
except:
  from TTS.utils.audio import AudioProcessor


from TTS.tts.models import setup_model
from TTS.config import load_config
from TTS.tts.models.vits import *

from TTS.tts.utils.speakers import SpeakerManager
from pydub import AudioSegment

# from google.colab import files
import librosa

from scipy.io.wavfile import write, read

import subprocess

'''
from google.colab import drive
drive.mount('/content/drive')

src_path = os.path.join(os.path.join(os.path.join(os.path.join(os.getcwd(), 'drive'), 'MyDrive'), 'Colab Notebooks'), 'best_model_latest.pth.tar')
dst_path = os.path.join(os.getcwd(), 'best_model.pth.tar')

shutil.copy(src_path, dst_path)
'''

TTS_PATH = "TTS/"

# add libraries into environment
sys.path.append(TTS_PATH) # set this if TTS is not installed globally

# Paths definition

OUT_PATH = 'out/'

# create output path
os.makedirs(OUT_PATH, exist_ok=True)

# model vars 
MODEL_PATH = 'best_model.pth.tar'
CONFIG_PATH = 'config.json'
TTS_LANGUAGES = "language_ids.json"
TTS_SPEAKERS = "speakers.json"
USE_CUDA = torch.cuda.is_available()

# load the config
C = load_config(CONFIG_PATH)

# load the audio processor
ap = AudioProcessor(**C.audio)

speaker_embedding = None

C.model_args['d_vector_file'] = TTS_SPEAKERS
C.model_args['use_speaker_encoder_as_loss'] = False

model = setup_model(C)
model.language_manager.set_language_ids_from_file(TTS_LANGUAGES)
# print(model.language_manager.num_languages, model.embedded_language_dim)
# print(model.emb_l)
cp = torch.load(MODEL_PATH, map_location=torch.device('cpu'))
# remove speaker encoder
model_weights = cp['model'].copy()
for key in list(model_weights.keys()):
  if "speaker_encoder" in key:
    del model_weights[key]

model.load_state_dict(model_weights)

model.eval()

if USE_CUDA:
    model = model.cuda()

# synthesize voice
use_griffin_lim = False

# Paths definition

CONFIG_SE_PATH = "config_se.json"
CHECKPOINT_SE_PATH = "SE_checkpoint.pth.tar"

# Load the Speaker encoder

SE_speaker_manager = SpeakerManager(encoder_model_path=CHECKPOINT_SE_PATH, encoder_config_path=CONFIG_SE_PATH, use_cuda=USE_CUDA)

# Define helper function

def compute_spec(ref_file):
  y, sr = librosa.load(ref_file, sr=ap.sample_rate)
  spec = ap.spectrogram(y)
  spec = torch.FloatTensor(spec).unsqueeze(0)
  return spec


def voice_conversion(ta, ra, da):

  target_audio = 'target.wav'
  reference_audio = 'reference.wav'
  driving_audio = 'driving.wav'

  write(target_audio, ta[0], ta[1])
  write(reference_audio, ra[0], ra[1])
  write(driving_audio, da[0], da[1])          

  # !ffmpeg-normalize $target_audio -nt rms -t=-27 -o $target_audio -ar 16000 -f
  # !ffmpeg-normalize $reference_audio -nt rms -t=-27 -o $reference_audio -ar 16000 -f
  # !ffmpeg-normalize $driving_audio -nt rms -t=-27 -o $driving_audio -ar 16000 -f

  files = [target_audio, reference_audio, driving_audio]

  for file in files:
      subprocess.run(["ffmpeg-normalize", file, "-nt", "rms", "-t=-27", "-o", file, "-ar", "16000", "-f"])

  # ta_ = read(target_audio)

  target_emb = SE_speaker_manager.compute_d_vector_from_clip([target_audio])
  target_emb = torch.FloatTensor(target_emb).unsqueeze(0)

  driving_emb = SE_speaker_manager.compute_d_vector_from_clip([reference_audio])
  driving_emb = torch.FloatTensor(driving_emb).unsqueeze(0)

  # Convert the voice

  driving_spec = compute_spec(driving_audio)
  y_lengths = torch.tensor([driving_spec.size(-1)])
  if USE_CUDA:
      ref_wav_voc, _, _ = model.voice_conversion(driving_spec.cuda(), y_lengths.cuda(), driving_emb.cuda(), target_emb.cuda())
      ref_wav_voc = ref_wav_voc.squeeze().cpu().detach().numpy()
  else:
      ref_wav_voc, _, _ = model.voice_conversion(driving_spec, y_lengths, driving_emb, target_emb)
      ref_wav_voc = ref_wav_voc.squeeze().detach().numpy()

  # print("Reference Audio after decoder:")
  # IPython.display.display(Audio(ref_wav_voc, rate=ap.sample_rate))

  return (ap.sample_rate, ref_wav_voc)

c3 = gr.Interface(
    fn=voice_conversion,
    inputs=[gr.Audio(label='Target Speaker - Reference Clip'), gr.Audio(label='Input Speaker - Reference Clip'), gr.Audio(label='Input Speaker - Clip To Convert')],
    outputs=gr.Audio(label='Target Speaker - Converted Clip'),
    examples=[['ntr.wav', 'timcast1.wav', 'timcast1.wav']], 
)

c1_m2 = gr.Interface(
    fn=voice_conversion,
    inputs=[gr.Audio(label='Target Speaker - Reference Clip'), gr.Audio(label='Input Speaker - Reference Clip', source='microphone'), gr.Audio(label='Input Speaker - Clip To Convert', source='microphone')],
    outputs=gr.Audio(label='Target Speaker - Converted Clip')
)

demo = gr.TabbedInterface([c3, c1_m2], ["Pre-Recorded", "Microphone"])

demo.launch(debug='True')