Method forward

modules/parallel_wavegan/models/source.py:158–202

Pulse train generator pulse_train, uv = forward(f0) input F0: tensor(batchsize=1, length, dim=1) f0 for unvoiced steps should be 0 output pulse_train: tensor(batchsize=1, length, dim) output uv: tensor(batchsize=1, length, 1) Note: self.l_s

(self, f0)

                                 flag_for_pulse=True)

    def forward(self, f0):
        """ Pulse train generator
        pulse_train, uv = forward(f0)
        input F0: tensor(batchsize=1, length, dim=1)
                  f0 for unvoiced steps should be 0
        output pulse_train: tensor(batchsize=1, length, dim)
        output uv: tensor(batchsize=1, length, 1)

        Note: self.l_sine doesn't make sure that the initial phase of
        a voiced segment is np.pi, the first pulse in a voiced segment
        may not be at the first time step within a voiced segment
        """
        with torch.no_grad():
            sine_wav, uv, noise = self.l_sinegen(f0)

            # sine without additive noise
            pure_sine = sine_wav - noise

            # step t corresponds to a pulse if
            # sine[t] > sine[t+1] & sine[t] > sine[t-1]
            # & sine[t-1], sine[t+1], and sine[t] are voiced
            # or
            # sine[t] is voiced, sine[t-1] is unvoiced
            # we use torch.roll to simulate sine[t+1] and sine[t-1]
            sine_1 = torch.roll(pure_sine, shifts=1, dims=1)
            uv_1 = torch.roll(uv, shifts=1, dims=1)
            uv_1[:, 0, :] = 0
            sine_2 = torch.roll(pure_sine, shifts=-1, dims=1)
            uv_2 = torch.roll(uv, shifts=-1, dims=1)
            uv_2[:, -1, :] = 0

            loc = (pure_sine > sine_1) * (pure_sine > sine_2) \
                  * (uv_1 > 0) * (uv_2 > 0) * (uv > 0) \
                  + (uv_1 < 1) * (uv > 0)

            # pulse train without noise
            pulse_train = pure_sine * loc

            # additive noise to pulse train
            # note that noise from sinegen is zero in voiced regions
            pulse_noise = torch.randn_like(pure_sine) * self.noise_std

            # with additive noise on pulse, and unvoiced regions
            pulse_train += pulse_noise * loc + pulse_noise * (1 - uv)
        return pulse_train, sine_wav, uv, pulse_noise


class SignalsConv1d(torch.nn.Module):