File size: 5,390 Bytes
fc9c1be |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 |
from turtle import title
import gradio as gr
import git
import os
os.system('git clone https://github.com/Edresson/Coqui-TTS -b multilingual-torchaudio-SE TTS')
os.system('pip install -q -e TTS/')
os.system('pip install -q torchaudio==0.9.0')
import sys
TTS_PATH = "TTS/"
# add libraries into environment
sys.path.append(TTS_PATH) # set this if TTS is not installed globally
import os
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 *
OUT_PATH = 'out/'
# create output path
os.makedirs(OUT_PATH, exist_ok=True)
# model vars
MODEL_PATH = '/home/user/app/best_model_latest.pth.tar'
CONFIG_PATH = '/home/user/app/config.json'
TTS_LANGUAGES = "/home/user/app/language_ids.json"
TTS_SPEAKERS = "/home/user/app/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
os.system('pip install -q pydub ffmpeg-normalize')
CONFIG_SE_PATH = "config_se.json"
CHECKPOINT_SE_PATH = "SE_checkpoint.pth.tar"
from TTS.tts.utils.speakers import SpeakerManager
from pydub import AudioSegment
import librosa
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
def greet(Text,Voicetoclone,VoiceMicrophone):
text= "%s" % (Text)
if Voicetoclone is not None:
reference_files= "%s" % (Voicetoclone)
print("path url")
print(Voicetoclone)
sample= str(Voicetoclone)
else:
reference_files= "%s" % (VoiceMicrophone)
print("path url")
print(VoiceMicrophone)
sample= str(VoiceMicrophone)
size= len(reference_files)*sys.getsizeof(reference_files)
size2= size / 1000000
if (size2 > 0.012) or len(text)>2000:
message="File is greater than 30mb or Text inserted is longer than 2000 characters. Please re-try with smaller sizes."
print(message)
raise SystemExit("File is greater than 30mb. Please re-try or Text inserted is longer than 2000 characters. Please re-try with smaller sizes.")
else:
os.system('ffmpeg-normalize $sample -nt rms -t=-27 -o $sample -ar 16000 -f')
reference_emb = SE_speaker_manager.compute_d_vector_from_clip(reference_files)
model.length_scale = 1 # scaler for the duration predictor. The larger it is, the slower the speech.
model.inference_noise_scale = 0.3 # defines the noise variance applied to the random z vector at inference.
model.inference_noise_scale_dp = 0.3 # defines the noise variance applied to the duration predictor z vector at inference.
text = text
model.language_manager.language_id_mapping
language_id = 0
print(" > text: {}".format(text))
wav, alignment, _, _ = synthesis(
model,
text,
C,
"cuda" in str(next(model.parameters()).device),
ap,
speaker_id=None,
d_vector=reference_emb,
style_wav=None,
language_id=language_id,
enable_eos_bos_chars=C.enable_eos_bos_chars,
use_griffin_lim=True,
do_trim_silence=False,
).values()
print("Generated Audio")
IPython.display.display(Audio(wav, rate=ap.sample_rate))
#file_name = text.replace(" ", "_")
#file_name = file_name.translate(str.maketrans('', '', string.punctuation.replace('_', ''))) + '.wav'
file_name="Audio.wav"
out_path = os.path.join(OUT_PATH, file_name)
print(" > Saving output to {}".format(out_path))
ap.save_wav(wav, out_path)
return out_path
demo = gr.Interface(
fn=greet,
inputs=[gr.inputs.Textbox(label='What would you like the voice to say? (max. 2000 characters per request)'),gr.Audio(type="filepath", source="upload",label='Please upload a voice to clone (max. 30mb)'),gr.Audio(source="microphone", type="filepath", streaming=True)],
outputs="audio",
title="Bilal's Voice Cloning Tool"
)
demo.launch() |