Griffin-Lim.py

self.power = 1.5
self.max_norm = 1.0
self.preemphasis = 0.97
self.ref_level_db = 20
self.min_level_db = -100

"""
Need to implement:

>>> np.clip
>>> np.power
>>> signal.lfilter
"""


def inv_spectrogram(self, spectrogram):
    """Converts spectrogram to waveform using librosa"""
    S = self._denormalize(spectrogram)
    S = self._db_to_amp(S + self.ref_level_db)  # Convert back to linear
    # Reconstruct phase
    return self.apply_inv_preemphasis(self._griffin_lim(S**self.power))
        
        
def _denormalize(self, S):
"""denormalize values"""
    S_denorm = S
    S_denorm = np.clip(S_denorm, 0, self.max_norm)
    S_denorm = (S_denorm * -self.min_level_db / self.max_norm) + self.min_level_db
     return S_denorm
    
    
def _db_to_amp(self, x):
    return np.power(10.0, x * 0.05)

# 对经过Griffin Lim合成出的audio wav, 使用preemphasis.
def apply_inv_preemphasis(self, x):
    return signal.lfilter([1], [1, -self.preemphasis], x)
 
########### 这个C++里已有实现 ############
def _griffin_lim(self, S):
    angles = np.exp(2j * np.pi * np.random.rand(*S.shape))
    S_complex = np.abs(S).astype(np.complex)
    y = self._istft(S_complex * angles)
    for i in range(self.griffin_lim_iters):
        angles = np.exp(1j * np.angle(self._stft(y)))
        y = self._istft(S_complex * angles)
    return y