Skip to content

hifigan

HIFIGAN

Bases: Module

Source code in speechain/module/vocoder/hifigan.py 
 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
class HIFIGAN(nn.Module):
    def __init__(
        self,
        in_channels=80,
        upsample_initial_channel=512,
        upsample_rates=(8, 8, 2, 2),
        upsample_kernel_sizes=(16, 16, 4, 4),
    ):
        super().__init__()

        self.pre_net = nn.Conv1d(in_channels, upsample_initial_channel, 7, padding=3)

        ups_in_channels = [
            upsample_initial_channel // (2**i) for i in range(len(upsample_rates))
        ]
        ups_out_channels = [
            upsample_initial_channel // (2 ** (i + 1))
            for i in range(len(upsample_rates))
        ]

        self.upsamples = nn.ModuleList()
        self.mrfs = nn.ModuleList()

        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
            self.upsamples.append(
                nn.ConvTranspose1d(
                    ups_in_channels[i],
                    ups_out_channels[i],
                    k,
                    stride=u,
                    padding=(k - u) // 2,
                )
            )
            self.mrfs.append(MRF(ups_out_channels[i]))

        self.post_net = nn.Sequential(
            nn.LeakyReLU(0.1),
            nn.Conv1d(ups_out_channels[-1], 1, 7, padding=3),
            nn.Tanh(),
        )

    def forward(self, x):
        x = self.pre_net(x)
        for up, mrf in zip(self.upsamples, self.mrfs):
            x = F.leaky_relu(x, 0.1)
            x = up(x)
            x = mrf(x)
        x = self.post_net(x)
        return x.squeeze(1)

    def decode_batch(self, feats):
        """Convert mel-spectrograms to waveforms
        Args:
            feats: (batch, time, n_mels)
        Returns:
            waveforms: (batch, time')
        """
        self.eval()
        with torch.no_grad():
            x = feats.transpose(1, 2)  # (B, n_mels, T)
            return self.forward(x)

    @classmethod
    def from_hparams(cls, source, savedir=None, run_opts=None):
        """Load pretrained model
        Args:
            source: Model identifier (e.g. "speechbrain/tts-hifigan-ljspeech")
            savedir: Directory to save model weights
            run_opts: Runtime options including device
        Returns:
            model: Loaded HiFiGAN model
        """
        model = cls()

        # Set device
        if run_opts and "device" in run_opts:
            model = model.to(run_opts["device"])

        # Load pretrained weights based on source
        if savedir:
            weights_path = os.path.join(savedir, "generator.pth")
            if os.path.exists(weights_path):
                model.load_state_dict(
                    torch.load(weights_path, map_location=run_opts["device"])
                )

        return model

decode_batch(feats)

Convert mel-spectrograms to waveforms Args: feats: (batch, time, n_mels) Returns: waveforms: (batch, time')

Source code in speechain/module/vocoder/hifigan.py 
 98
 99
100
101
102
103
104
105
106
107
108
def decode_batch(self, feats):
    """Convert mel-spectrograms to waveforms
    Args:
        feats: (batch, time, n_mels)
    Returns:
        waveforms: (batch, time')
    """
    self.eval()
    with torch.no_grad():
        x = feats.transpose(1, 2)  # (B, n_mels, T)
        return self.forward(x)

from_hparams(source, savedir=None, run_opts=None) classmethod

Load pretrained model Args: source: Model identifier (e.g. "speechbrain/tts-hifigan-ljspeech") savedir: Directory to save model weights run_opts: Runtime options including device Returns: model: Loaded HiFiGAN model

Source code in speechain/module/vocoder/hifigan.py 
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
@classmethod
def from_hparams(cls, source, savedir=None, run_opts=None):
    """Load pretrained model
    Args:
        source: Model identifier (e.g. "speechbrain/tts-hifigan-ljspeech")
        savedir: Directory to save model weights
        run_opts: Runtime options including device
    Returns:
        model: Loaded HiFiGAN model
    """
    model = cls()

    # Set device
    if run_opts and "device" in run_opts:
        model = model.to(run_opts["device"])

    # Load pretrained weights based on source
    if savedir:
        weights_path = os.path.join(savedir, "generator.pth")
        if os.path.exists(weights_path):
            model.load_state_dict(
                torch.load(weights_path, map_location=run_opts["device"])
            )

    return model