File size: 4,224 Bytes
9016314
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88b5ee2
9016314
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88b5ee2
 
 
 
 
 
 
 
9016314
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8e45d4b
9016314
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5d4265b
9016314
 
 
 
 
 
 
 
 
 
 
5341d30
9016314
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import time

import numpy as np
import torch
import soundfile as sf
import argparse

from SVCNN import SVCNN
from utils.tools import extract_voiced_area
from utils.extract_pitch import extract_pitch_ref as extract_pitch, coarse_f0

SPEAKER_INFORMATION_WEIGHTS = [
    0, 0, 0, 0, 0, 0,  # layer 0-5
    1.0, 0, 0, 0,
    0, 0, 0, 0, 0, 0,  # layer 15
    0, 0, 0, 0, 0, 0,  # layer 16-21
    0,  # layer 22
    0, 0  # layer 23-24
]
SPEAKER_INFORMATION_LAYER = 6


APPLIED_INFORMATION_WEIGHTS = [
    0, 0, 0, 0, 0, 0,  # layer 0-5
    0, 0, 0, 0,
    0, 0, 0, 0, 0, 0,  # layer 15
    0, 0, 0, 0, 0.2, 0.2,  # layer 16-21
    0.2,  # layer 22
    0.2, 0.2  # layer 23-24
]


def svc(model, src_wav_path, ref_wav_path, synth_set_path=None, f0_factor=0., speech_enroll=False, out_dir="output", hallucinated_set_path=None, device='cpu'):
    
    wav_name = os.path.basename(src_wav_path).split('.')[0]
    ref_name = os.path.basename(ref_wav_path).split('.')[0]

    f0_src, f0_factor = extract_pitch(src_wav_path, ref_wav_path, predefined_factor=f0_factor, speech_enroll=speech_enroll)

    pitch_src = coarse_f0(f0_src)

    query_mask = extract_voiced_area(src_wav_path, hop_size=480, energy_thres=0.1)
    query_mask = torch.from_numpy(query_mask).to(device)

    synth_weights = torch.tensor(
        SPEAKER_INFORMATION_WEIGHTS, device=device)[:, None]
    query_seq = model.get_features(
        src_wav_path, weights=synth_weights)

    if synth_set_path:
        synth_set = torch.load(synth_set_path).to(device)
    else:
        synth_set = model.get_matching_set(ref_wav_path).to(device)

    if hallucinated_set_path:
        hallucinated_set = torch.from_numpy(np.load(hallucinated_set_path)).to(device)
        synth_set = torch.cat([synth_set, hallucinated_set], dim=0)
    
    query_len = query_seq.shape[0]
    if len(query_mask) > query_len:
        query_mask = query_mask[:query_len]
    else:
        p = query_len - len(query_mask)
        query_mask = np.pad(query_mask, (0, p))

    f0_len = query_len*2
    if len(f0_src) > f0_len:
        f0_src = f0_src[:f0_len]
        pitch_src = pitch_src[:f0_len]
    else:
        p = f0_len-len(f0_src)
        f0_src = np.pad(f0_src, (0, p), mode='edge')
        pitch_src = np.pad(pitch_src, (0, p), mode='edge')
    
    print(query_seq.shape)
    print(synth_set.shape)

    f0_src = torch.from_numpy(f0_src).float().to(device)
    pitch_src = torch.from_numpy(pitch_src).to(device)

    out_wav = model.match(query_seq, f0_src, pitch_src, synth_set, topk=4, query_mask=query_mask)
    # out_wav is (T,) tensor converted 16kHz output wav using k=4 for kNN.
    os.makedirs(out_dir, exist_ok=True)
    wfname = f'{out_dir}/NeuCoSVCv2.wav'

    sf.write(wfname, out_wav.numpy(), 24000)


def main(a):
    model_ckpt_path = a.model_ckpt_path
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    # device = 'cpu'
    print(f'using {device} for inference')

    f0factor = pow(2, a.key_shift / 12) if a.key_shift else 0.

    speech_enroll = a.speech_enroll
    model = SVCNN(model_ckpt_path, device=device)

    t0 = time.time()
    svc(model, a.src_wav_path, a.ref_wav_path, out_dir=a.out_dir, device=device, f0_factor=f0factor, speech_enroll=speech_enroll)
    t1 = time.time()
    print(f"{t1-t0:.2f}s to perfrom the conversion")


if __name__ == '__main__':
    parser = argparse.ArgumentParser()

    parser.add_argument('--src_wav_path', required=True)
    parser.add_argument('--ref_wav_path', required=True)
    parser.add_argument('--model_ckpt_path',
                        default='ckpt/G_150k.pt')
    parser.add_argument('--out_dir', default='output_svc')
    parser.add_argument(
        '--key_shift', type=int,
        help='Adjust the pitch of the source singing. Tone the song up or down in semitones.'
    )
    parser.add_argument(
        '--speech_enroll', action='store_true',
        help='When using speech as the reference audio, the pitch of the reference audio will be increased by 1.2 times \
            when performing pitch shift to cover the pitch gap between singing and speech. \
            Note: This option is invalid when key_shift is specified.'
    )

    a = parser.parse_args()

    main(a)