forked from PlayVoice/whisper-vits-svc
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsvc_inference.py
167 lines (141 loc) · 5.77 KB
/
svc_inference.py
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
166
167
import os
import torch
import argparse
import numpy as np
from omegaconf import OmegaConf
from scipy.io.wavfile import write
from vits.models import SynthesizerInfer
from pitch import load_csv_pitch
def load_svc_model(checkpoint_path, model):
assert os.path.isfile(checkpoint_path)
checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
saved_state_dict = checkpoint_dict["model_g"]
state_dict = model.state_dict()
new_state_dict = {}
for k, v in state_dict.items():
new_state_dict[k] = saved_state_dict[k]
model.load_state_dict(new_state_dict)
return model
def main(args):
if (args.ppg == None):
args.ppg = "svc_tmp.ppg.npy"
print(
f"Auto run : python whisper/inference.py -w {args.wave} -p {args.ppg}")
os.system(f"python whisper/inference.py -w {args.wave} -p {args.ppg}")
if (args.pit == None):
args.pit = "svc_tmp.pit.csv"
print(
f"Auto run : python pitch/inference.py -w {args.wave} -p {args.pit}")
os.system(f"python pitch/inference.py -w {args.wave} -p {args.pit}")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
hp = OmegaConf.load(args.config)
model = SynthesizerInfer(
hp.data.filter_length // 2 + 1,
hp.data.segment_size // hp.data.hop_length,
hp)
load_svc_model(args.model, model)
model.eval()
model.to(device)
spk = np.load(args.spk)
spk = torch.FloatTensor(spk)
ppg = np.load(args.ppg)
ppg = np.repeat(ppg, 2, 0) # 320 PPG -> 160 * 2
ppg = torch.FloatTensor(ppg)
pit = load_csv_pitch(args.pit)
if (args.statics == None):
print("don't use pitch shift")
else:
source = pit[pit > 0]
source_ave = source.mean()
source_min = source.min()
source_max = source.max()
print(f"source pitch statics: mean={source_ave:0.1f}, \
min={source_min:0.1f}, max={source_max:0.1f}")
singer_ave, singer_min, singer_max = np.load(args.statics)
print(f"singer pitch statics: mean={singer_ave:0.1f}, \
min={singer_min:0.1f}, max={singer_max:0.1f}")
shift = np.log2(singer_ave/source_ave) * 12
if (singer_ave >= source_ave):
shift = np.floor(shift)
else:
shift = np.ceil(shift)
shift = 2 ** (shift / 12)
pit = pit * shift
pit = torch.FloatTensor(pit)
len_pit = pit.size()[0]
len_ppg = ppg.size()[0]
len_min = min(len_pit, len_ppg)
pit = pit[:len_min]
ppg = ppg[:len_min, :]
with torch.no_grad():
spk = spk.unsqueeze(0).to(device)
source = pit.unsqueeze(0).to(device)
source = model.pitch2source(source)
pitwav = model.source2wav(source)
write("svc_out_pit.wav", hp.data.sampling_rate, pitwav)
hop_size = hp.data.hop_length
all_frame = len_min
hop_frame = 10
out_chunk = 2500 # 25 S
out_index = 0
out_audio = []
has_audio = False
while (out_index + out_chunk < all_frame):
has_audio = True
if (out_index == 0): # start frame
cut_s = 0
cut_s_out = 0
else:
cut_s = out_index - hop_frame
cut_s_out = hop_frame * hop_size
if (out_index + out_chunk + hop_frame > all_frame): # end frame
cut_e = out_index + out_chunk
cut_e_out = 0
else:
cut_e = out_index + out_chunk + hop_frame
cut_e_out = -1 * hop_frame * hop_size
sub_ppg = ppg[cut_s:cut_e, :].unsqueeze(0).to(device)
sub_pit = pit[cut_s:cut_e].unsqueeze(0).to(device)
sub_len = torch.LongTensor([cut_e - cut_s]).to(device)
sub_har = source[:, :, cut_s *
hop_size:cut_e * hop_size].to(device)
sub_out = model.inference(sub_ppg, sub_pit, spk, sub_len, sub_har)
sub_out = sub_out[0, 0].data.cpu().detach().numpy()
sub_out = sub_out[cut_s_out:cut_e_out]
out_audio.extend(sub_out)
out_index = out_index + out_chunk
if (out_index < all_frame):
if (has_audio):
cut_s = out_index - hop_frame
cut_s_out = hop_frame * hop_size
else:
cut_s = 0
cut_s_out = 0
sub_ppg = ppg[cut_s:, :].unsqueeze(0).to(device)
sub_pit = pit[cut_s:].unsqueeze(0).to(device)
sub_len = torch.LongTensor([all_frame - cut_s]).to(device)
sub_har = source[:, :, cut_s * hop_size:].to(device)
sub_out = model.inference(sub_ppg, sub_pit, spk, sub_len, sub_har)
sub_out = sub_out[0, 0].data.cpu().detach().numpy()
sub_out = sub_out[cut_s_out:]
out_audio.extend(sub_out)
out_audio = np.asarray(out_audio)
write("svc_out.wav", hp.data.sampling_rate, out_audio)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config', type=str, required=True,
help="yaml file for config.")
parser.add_argument('--model', type=str, required=True,
help="path of model for evaluation")
parser.add_argument('--wave', type=str, required=True,
help="Path of raw audio.")
parser.add_argument('--spk', type=str, required=True,
help="Path of speaker.")
parser.add_argument('--ppg', type=str,
help="Path of content vector.")
parser.add_argument('--pit', type=str,
help="Path of pitch csv file.")
parser.add_argument('--statics', type=str,
help="Path of pitch statics.")
args = parser.parse_args()
main(args)