Book a Demo!
CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutPoliciesSign UpSign In
prophesier
GitHub Repository: prophesier/diff-svc
 Path: blob/main/network/vocoders/nsf_hifigan.py
701 views
1
import os

2
import torch

3
from modules.nsf_hifigan.models import load_model, Generator

4
from modules.nsf_hifigan.nvSTFT import load_wav_to_torch, STFT

5
from utils.hparams import hparams

6
from network.vocoders.base_vocoder import BaseVocoder, register_vocoder

7


8
@register_vocoder

9
class NsfHifiGAN(BaseVocoder):

10
    def __init__(self, device=None):

11
        if device is None:

12
            device = 'cuda' if torch.cuda.is_available() else 'cpu'

13
        self.device = device

14
        model_path = hparams['vocoder_ckpt']

15
        if os.path.exists(model_path):

16
            print('| Load HifiGAN: ', model_path)

17
            self.model, self.h = load_model(model_path, device=self.device)

18
        else:

19
            print('Error: HifiGAN model file is not found!')

20


21
    def spec2wav_torch(self, mel, **kwargs): # mel: [B, T, bins]

22
        if self.h.sampling_rate != hparams['audio_sample_rate']:

23
            print('Mismatch parameters: hparams[\'audio_sample_rate\']=',hparams['audio_sample_rate'],'!=',self.h.sampling_rate,'(vocoder)')

24
        if self.h.num_mels != hparams['audio_num_mel_bins']:

25
            print('Mismatch parameters: hparams[\'audio_num_mel_bins\']=',hparams['audio_num_mel_bins'],'!=',self.h.num_mels,'(vocoder)')

26
        if self.h.n_fft != hparams['fft_size']:

27
            print('Mismatch parameters: hparams[\'fft_size\']=',hparams['fft_size'],'!=',self.h.n_fft,'(vocoder)')

28
        if self.h.win_size != hparams['win_size']:

29
            print('Mismatch parameters: hparams[\'win_size\']=',hparams['win_size'],'!=',self.h.win_size,'(vocoder)')

30
        if self.h.hop_size != hparams['hop_size']:

31
            print('Mismatch parameters: hparams[\'hop_size\']=',hparams['hop_size'],'!=',self.h.hop_size,'(vocoder)')

32
        if self.h.fmin != hparams['fmin']:

33
            print('Mismatch parameters: hparams[\'fmin\']=',hparams['fmin'],'!=',self.h.fmin,'(vocoder)')

34
        if self.h.fmax != hparams['fmax']:

35
            print('Mismatch parameters: hparams[\'fmax\']=',hparams['fmax'],'!=',self.h.fmax,'(vocoder)')

36
        with torch.no_grad():

37
            c = mel.transpose(2, 1) #[B, T, bins]

38
            #log10 to log mel

39
            c = 2.30259 * c

40
            f0 = kwargs.get('f0') #[B, T]

41
            if f0 is not None and hparams.get('use_nsf'):

42
                y = self.model(c, f0).view(-1)

43
            else:

44
                y = self.model(c).view(-1)

45
        return y

46


47
    def spec2wav(self, mel, **kwargs):

48
        if self.h.sampling_rate != hparams['audio_sample_rate']:

49
            print('Mismatch parameters: hparams[\'audio_sample_rate\']=',hparams['audio_sample_rate'],'!=',self.h.sampling_rate,'(vocoder)')

50
        if self.h.num_mels != hparams['audio_num_mel_bins']:

51
            print('Mismatch parameters: hparams[\'audio_num_mel_bins\']=',hparams['audio_num_mel_bins'],'!=',self.h.num_mels,'(vocoder)')

52
        if self.h.n_fft != hparams['fft_size']:

53
            print('Mismatch parameters: hparams[\'fft_size\']=',hparams['fft_size'],'!=',self.h.n_fft,'(vocoder)')

54
        if self.h.win_size != hparams['win_size']:

55
            print('Mismatch parameters: hparams[\'win_size\']=',hparams['win_size'],'!=',self.h.win_size,'(vocoder)')

56
        if self.h.hop_size != hparams['hop_size']:

57
            print('Mismatch parameters: hparams[\'hop_size\']=',hparams['hop_size'],'!=',self.h.hop_size,'(vocoder)')

58
        if self.h.fmin != hparams['fmin']:

59
            print('Mismatch parameters: hparams[\'fmin\']=',hparams['fmin'],'!=',self.h.fmin,'(vocoder)')

60
        if self.h.fmax != hparams['fmax']:

61
            print('Mismatch parameters: hparams[\'fmax\']=',hparams['fmax'],'!=',self.h.fmax,'(vocoder)')

62
        with torch.no_grad():

63
            c = torch.FloatTensor(mel).unsqueeze(0).transpose(2, 1).to(self.device)

64
            #log10 to log mel

65
            c = 2.30259 * c

66
            f0 = kwargs.get('f0')

67
            if f0 is not None and hparams.get('use_nsf'):

68
                f0 = torch.FloatTensor(f0[None, :]).to(self.device)

69
                y = self.model(c, f0).view(-1)

70
            else:

71
                y = self.model(c).view(-1)

72
        wav_out = y.cpu().numpy()

73
        return wav_out

74


75
    @staticmethod

76
    def wav2spec(inp_path, device=None):

77
        if device is None:

78
            device = 'cuda' if torch.cuda.is_available() else 'cpu'

79
        sampling_rate = hparams['audio_sample_rate']

80
        num_mels = hparams['audio_num_mel_bins']

81
        n_fft = hparams['fft_size']

82
        win_size =hparams['win_size']

83
        hop_size = hparams['hop_size']

84
        fmin = hparams['fmin']

85
        fmax = hparams['fmax']

86
        stft = STFT(sampling_rate, num_mels, n_fft, win_size, hop_size, fmin, fmax)

87
        with torch.no_grad():

88
            wav_torch, _ = load_wav_to_torch(inp_path, target_sr=stft.target_sr)

89
            mel_torch = stft.get_mel(wav_torch.unsqueeze(0).to(device)).squeeze(0).T

90
            #log mel to log10 mel

91
            mel_torch = 0.434294 * mel_torch

92
            return wav_torch.cpu().numpy(), mel_torch.cpu().numpy()

93