def __convert_single_only_au_AutoVC_format_to_dataset__(
self, filename, build_train_dataset=True):
"""
Convert a single file (only audio in AutoVC embedding format) to numpy arrays
:param filename:
:param is_map_to_std_face:
:return:
"""
global_clip_index, video_name = filename
# audio_file = os.path.join(self.src_dir, 'raw_wav', '{}.wav'.
# format(video_name[:-4]))
audio_file = os.path.join(
self.src_dir, 'raw_wav',
'{:05d}_{}_audio.wav'.format(global_clip_index, video_name[:-4]))
if (not build_train_dataset):
import shutil
audio_file = os.path.join(
self.src_dir, 'raw_wav',
'{:05d}_{}_audio.wav'.format(global_clip_index,
video_name[:-4]))
shutil.copy(
os.path.join(self.src_dir, 'test_wav_files', video_name),
audio_file)
sound = AudioSegment.from_file(audio_file, "wav")
normalized_sound = match_target_amplitude(sound, -20.0)
normalized_sound.export(audio_file, format='wav')
from src.autovc.retrain_version.vocoder_spec.extract_f0_func import extract_f0_func_audiofile
S, f0_norm = extract_f0_func_audiofile(audio_file, 'M')
from src.autovc.utils import quantize_f0_interp
f0_onehot = quantize_f0_interp(f0_norm)
from thirdparty.resemblyer_util.speaker_emb import get_spk_emb
mean_emb, _ = get_spk_emb(audio_file)
return S, mean_emb, f0_onehot
'examples_cartoon/{}_face_close_mouth.txt'.format(DEMO_CH))
''' STEP 3: Generate audio data as input to audio branch '''
au_data = []
au_emb = []
ains = glob.glob1('examples', '*.wav')
ains = [item for item in ains if item is not 'tmp.wav']
ains.sort()
for ain in ains:
os.system(
'ffmpeg -y -loglevel error -i examples/{} -ar 16000 examples/tmp.wav'.
format(ain))
shutil.copyfile('examples/tmp.wav', 'examples/{}'.format(ain))
# au embedding
from thirdparty.resemblyer_util.speaker_emb import get_spk_emb
me, ae = get_spk_emb('examples/{}'.format(ain))
au_emb.append(me.reshape(-1))
print('Processing audio file', ain)
c = AutoVC_mel_Convertor('examples')
au_data_i = c.convert_single_wav_to_autovc_input(
audio_filename=os.path.join('examples', ain),
autovc_model_path=opt_parser.load_AUTOVC_name)
au_data += au_data_i
# os.remove(os.path.join('examples', 'tmp.wav'))
if (os.path.isfile('examples/tmp.wav')):
os.remove('examples/tmp.wav')
fl_data = []
rot_tran, rot_quat, anchor_t_shape = [], [], []
for au, info in au_data:
def convert_single_wav_to_autovc_input(self, audio_filename,
autovc_model_path):
def pad_seq(x, base=32):
len_out = int(base * ceil(float(x.shape[0]) / base))
len_pad = len_out - x.shape[0]
assert len_pad >= 0
return np.pad(x, ((0, len_pad), (0, 0)), 'constant'), len_pad
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
G = Generator(16, 256, 512, 16).eval().to(device)
g_checkpoint = torch.load(autovc_model_path, map_location=device)
G.load_state_dict(g_checkpoint['model'])
emb = np.loadtxt('src/autovc/retrain_version/obama_emb.txt')
emb_trg = torch.from_numpy(
emb[np.newaxis, :].astype('float32')).to(device)
aus = []
audio_file = audio_filename
sound = AudioSegment.from_file(audio_file, "wav")
normalized_sound = match_target_amplitude(sound, -20.0)
normalized_sound.export(audio_file, format='wav')
from src.autovc.retrain_version.vocoder_spec.extract_f0_func import extract_f0_func_audiofile
x_real_src, f0_norm = extract_f0_func_audiofile(audio_file, 'F')
from src.autovc.utils import quantize_f0_interp
f0_org_src = quantize_f0_interp(f0_norm)
from thirdparty.resemblyer_util.speaker_emb import get_spk_emb
emb, _ = get_spk_emb(audio_file)
''' normal length version '''
# x_real, len_pad = pad_seq(x_real_src.astype('float32'))
# f0_org, _ = pad_seq(f0_org_src.astype('float32'))
# x_real = torch.from_numpy(x_real[np.newaxis, :].astype('float32')).to(device)
# emb_org = torch.from_numpy(emb[np.newaxis, :].astype('float32')).to(device)
# f0_org = torch.from_numpy(f0_org[np.newaxis, :].astype('float32')).to(device)
# print('source shape:', x_real.shape, emb_org.shape, emb_trg.shape, f0_org.shape)
#
# with torch.no_grad():
# x_identic, x_identic_psnt, code_real = G(x_real, emb_org, f0_org, emb_trg, f0_org)
# print('converted shape:', x_identic_psnt.shape, code_real.shape)
''' long split version '''
l = x_real_src.shape[0]
x_identic_psnt = []
step = 4096
for i in range(0, l, step):
x_real = x_real_src[i:i + step]
f0_org = f0_org_src[i:i + step]
x_real, len_pad = pad_seq(x_real.astype('float32'))
f0_org, _ = pad_seq(f0_org.astype('float32'))
x_real = torch.from_numpy(
x_real[np.newaxis, :].astype('float32')).to(device)
emb_org = torch.from_numpy(
emb[np.newaxis, :].astype('float32')).to(device)
# emb_trg = torch.from_numpy(emb[np.newaxis, :].astype('float32')).to(device)
f0_org = torch.from_numpy(
f0_org[np.newaxis, :].astype('float32')).to(device)
print('source shape:', x_real.shape, emb_org.shape, emb_trg.shape,
f0_org.shape)
with torch.no_grad():
x_identic, x_identic_psnt_i, code_real = G(
x_real, emb_org, f0_org, emb_trg, f0_org)
x_identic_psnt.append(x_identic_psnt_i)
x_identic_psnt = torch.cat(x_identic_psnt, dim=1)
print('converted shape:', x_identic_psnt.shape, code_real.shape)
if len_pad == 0:
uttr_trg = x_identic_psnt[0, :, :].cpu().numpy()
else:
uttr_trg = x_identic_psnt[0, :-len_pad, :].cpu().numpy()
aus.append((uttr_trg, (0, audio_filename, emb)))
return aus