File size: 3,370 Bytes
85ce65e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from torchaudio.transforms import Resample

from vdecoder.nsf_hifigan.models import load_config, load_model
from vdecoder.nsf_hifigan.nvSTFT import STFT


class Vocoder:
    def __init__(self, vocoder_type, vocoder_ckpt, device = None):
        if device is None:
            device = 'cuda' if torch.cuda.is_available() else 'cpu'
        self.device = device
        
        if vocoder_type == 'nsf-hifigan':
            self.vocoder = NsfHifiGAN(vocoder_ckpt, device = device)
        elif vocoder_type == 'nsf-hifigan-log10':
            self.vocoder = NsfHifiGANLog10(vocoder_ckpt, device = device)
        else:
            raise ValueError(f" [x] Unknown vocoder: {vocoder_type}")
            
        self.resample_kernel = {}
        self.vocoder_sample_rate = self.vocoder.sample_rate()
        self.vocoder_hop_size = self.vocoder.hop_size()
        self.dimension = self.vocoder.dimension()
        
    def extract(self, audio, sample_rate, keyshift=0):
                
        # resample
        if sample_rate == self.vocoder_sample_rate:
            audio_res = audio
        else:
            key_str = str(sample_rate)
            if key_str not in self.resample_kernel:
                self.resample_kernel[key_str] = Resample(sample_rate, self.vocoder_sample_rate, lowpass_filter_width = 128).to(self.device)
            audio_res = self.resample_kernel[key_str](audio)    
        
        # extract
        mel = self.vocoder.extract(audio_res, keyshift=keyshift) # B, n_frames, bins
        return mel
   
    def infer(self, mel, f0):
        f0 = f0[:,:mel.size(1),0] # B, n_frames
        audio = self.vocoder(mel, f0)
        return audio
        
        
class NsfHifiGAN(torch.nn.Module):
    def __init__(self, model_path, device=None):
        super().__init__()
        if device is None:
            device = 'cuda' if torch.cuda.is_available() else 'cpu'
        self.device = device
        self.model_path = model_path
        self.model = None
        self.h = load_config(model_path)
        self.stft = STFT(
                self.h.sampling_rate, 
                self.h.num_mels, 
                self.h.n_fft, 
                self.h.win_size, 
                self.h.hop_size, 
                self.h.fmin, 
                self.h.fmax)
    
    def sample_rate(self):
        return self.h.sampling_rate
        
    def hop_size(self):
        return self.h.hop_size
    
    def dimension(self):
        return self.h.num_mels
        
    def extract(self, audio, keyshift=0):       
        mel = self.stft.get_mel(audio, keyshift=keyshift).transpose(1, 2) # B, n_frames, bins
        return mel
    
    def forward(self, mel, f0):
        if self.model is None:
            print('| Load HifiGAN: ', self.model_path)
            self.model, self.h = load_model(self.model_path, device=self.device)
        with torch.no_grad():
            c = mel.transpose(1, 2)
            audio = self.model(c, f0)
            return audio

class NsfHifiGANLog10(NsfHifiGAN):    
    def forward(self, mel, f0):
        if self.model is None:
            print('| Load HifiGAN: ', self.model_path)
            self.model, self.h = load_model(self.model_path, device=self.device)
        with torch.no_grad():
            c = 0.434294 * mel.transpose(1, 2)
            audio = self.model(c, f0)
            return audio