shifter.py 文件源码

def _high_frequency_completion(self, x, transformed):
        """
        Please see Sect. 3.2 and 3.3 in the following paper to know why we complete the
        unvoiced synthesized voice of the original voice into high frequency range
        of F0 transformed voice.

        - K. Kobayashi et al., "F0 transformation techniques for statistical voice
        conversion with direct waveform modification with spectral differential,"
        Proc. IEEE SLT 2016, pp. 693-700. 2016.
        """
        # construct feature extractor and synthesis
        feat = FeatureExtractor(fs=self.fs)
        f0, spc, ap = feat.analyze(x)
        uf0 = np.zeros(len(f0))

        # synthesis
        synth = Synthesizer(fs=self.fs)
        unvoice_anasyn = synth.synthesis_spc(uf0, spc, ap)

        # HPF for synthesized speech
        fil = firwin(255, self.f0rate, pass_zero=False)
        HPFed_unvoice_anasyn = lfilter(fil, 1, unvoice_anasyn)

        if len(HPFed_unvoice_anasyn) > len(transformed):
            return transformed + HPFed_unvoice_anasyn[:len(transformed)]
        else:
            transformed[:len(HPFed_unvoice_anasyn)] += HPFed_unvoice_anasyn
            return transformed