app.py

'''
import gradio as gr

def greet(name):
    return "Hello " + name + "!!"

iface = gr.Interface(fn=greet, inputs="text", outputs="text")
iface.launch()
'''
import gradio
import os
import shutil
import gradio as gr
import sys
import string
import time
import argparse
import json
import numpy as np
import torch
import librosa
import subprocess

from pydub import AudioSegment
from scipy.io.wavfile import write, read
from transformers import WavLMModel

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

import utils
from models import SynthesizerTrn
from mel_processing import mel_spectrogram_torch
from speaker_encoder.voice_encoder import SpeakerEncoder

TTS_PATH = "TTS/"
sys.path.append(TTS_PATH) # set this if TTS is not installed globally

OUT_PATH = 'out/'
os.makedirs(OUT_PATH, exist_ok=True)

TTS_SPEAKERS = "yourtts_config/speakers.json"
USE_CUDA = torch.cuda.is_available()
device = torch.device("cuda" if USE_CUDA else "cpu")

CONFIG_PATH = 'yourtts_config/config.json'
C = load_config(CONFIG_PATH)
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)

TTS_LANGUAGES = "yourtts_config/language_ids.json"
model.language_manager.set_language_ids_from_file(TTS_LANGUAGES)

# print(model.language_manager.num_languages, model.embedded_language_dim)
# print(model.emb_l)

MODEL_PATH = 'yourtts_config/best_model.pth.tar'
cp = torch.load(MODEL_PATH, map_location=torch.device('cpu'))

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()

use_griffin_lim = False

CONFIG_SE_PATH = "yourtts_config/config_se.json"
CHECKPOINT_SE_PATH = "yourtts_config/SE_checkpoint.pth.tar"
SE_speaker_manager = SpeakerManager(encoder_model_path=CHECKPOINT_SE_PATH, encoder_config_path=CONFIG_SE_PATH, use_cuda=USE_CUDA)

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

print("Loading FreeVC...")
hps = utils.get_hparams_from_file("configs/freevc.json")
freevc = SynthesizerTrn(
    hps.data.filter_length // 2 + 1,
    hps.train.segment_size // hps.data.hop_length,
    **hps.model).to(device)
_ = freevc.eval()
_ = utils.load_checkpoint("checkpoints/freevc.pth", freevc, None)
smodel = SpeakerEncoder('speaker_encoder/ckpt/pretrained_bak_5805000.pt')

print("Loading WavLM for content...")
cmodel = utils.get_cmodel(device).to(device)
# cmodel = WavLMModel.from_pretrained("microsoft/wavlm-large").to(device)

def voice_conversion_yourtts(da, ta, normalize=False):

  # write(target_audio, ta[0], ta[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 = [da, ta]

  subprocess.run(["ffmpeg-normalize", da, "-nt", "rms", "-t=-27", "-o", "source_yourtts.wav", "-ar", "16000", "-f"])
  subprocess.run(["ffmpeg-normalize", ta, "-nt", "rms", "-t=-27", "-o", "target_yourtts.wav", "-ar", "16000", "-f"])  

  # ta_ = read(target_audio)

  target_emb = SE_speaker_manager.compute_d_vector_from_clip(["target_yourtts.wav"])
  target_emb = torch.FloatTensor(target_emb).unsqueeze(0)

  driving_emb = SE_speaker_manager.compute_d_vector_from_clip(["source_yourtts.wav"])
  driving_emb = torch.FloatTensor(driving_emb).unsqueeze(0)

  # Convert the voice

  driving_spec = compute_spec("source_yourtts.wav")
  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)


def voice_conversion_freevc(src, tgt, normalize=False):
    with torch.no_grad():
        subprocess.run(["ffmpeg-normalize", tgt, "-nt", "rms", "-t=-27", "-o", "target_fvc.wav", "-ar", "16000", "-f"])
        wav_tgt, _ = librosa.load("target_fvc.wav", sr=hps.data.sampling_rate)
        wav_tgt, _ = librosa.effects.trim(wav_tgt, top_db=20)
        g_tgt = smodel.embed_utterance(wav_tgt)
        g_tgt = torch.from_numpy(g_tgt).unsqueeze(0).to(device)
        subprocess.run(["ffmpeg-normalize", src, "-nt", "rms", "-t=-27", "-o", "source_fvc.wav", "-ar", "16000", "-f"])
        wav_src, _ = librosa.load("source_fvc.wav", sr=hps.data.sampling_rate)
        wav_src = torch.from_numpy(wav_src).unsqueeze(0).to(device)
        # c = cmodel(wav_src).last_hidden_state.transpose(1, 2).to(device)
        c = utils.get_content(cmodel, wav_src)
        audio = freevc.infer(c, g=g_tgt)
        audio = audio[0][0].data.cpu().float().numpy()
        write("out.wav", hps.data.sampling_rate, audio)
        out = "out.wav"
    return out

model1 = gr.Dropdown(choices=["FreeVC", "YourTTS"], value="FreeVC",type="value", label="Model") 
model2 = gr.Dropdown(choices=["FreeVC", "YourTTS"], value="FreeVC",type="value", label="Model") 

audio1 = gr.inputs.Audio(label="Source Speaker - Input Audio", type='filepath')
audio2 = gr.inputs.Audio(label="Target Speaker - Input Audio", type='filepath')
microphone = gr.inputs.Audio(label="Source Speaker - Input Audio", source='microphone')
audio3 = gr.inputs.Audio(label="Target Speaker - Input Audio", type='filepath')

inputs_1 = [model1, audio1, audio2]
inputs_2 = [model2, microphone, audio3]

outputs_1 =  gr.outputs.Audio(label="Target Speaker - Output Audio", type='filepath')
outputs_2 =  gr.outputs.Audio(label="Target Speaker - Output Audio", type='filepath')

def voice_conversion(mod, sa, ta):
    
    if mod=='FreeVC':
        return voice_conversion_freevc(sa, ta)
    else:
        return voice_conversion_yourtts(sa, ta)
    
examples_1 = [['FreeVC', 'sample_inputs/ntr.wav', 'sample_inputs/timcast1.wav'], ['YourTTS', 'sample_inputs/ntr.wav', 'sample_inputs/timcast1.wav']]

vc_1 = gr.Interface(
    fn=voice_conversion,
    inputs=inputs_1,
    outputs=outputs_1,
    examples=examples_1,
    description="Use this cool tool to convert your voice to another person's! \n Upload files in wav format for the source speaker and the target speaker.\n \nThis demonstration is made by T B Ramkamal, for partial credit towards completion of my Dual Degree Project"
)

vc_2 = gr.Interface(
    fn=voice_conversion,
    inputs=inputs_2,
    outputs=outputs_2,
    description="Use this cool tool to convert your voice to another person's! \n Upload files in wav format for the target speaker and record the voice of the input speaker using the microphone.\n \nThis demonstration is made by T B Ramkamal, for partial credit towards completion of my Dual Degree Project"
)

demo = gr.TabbedInterface([vc_1, vc_2], ["wav Input", "Microphone Input"], title="Voice Conversion Demo")
demo.launch(debug='True')