def __init__(self,
filter_length=1024,
hop_length=256, win_length=1024,
n_mel_channels=80,
# sampling_rate=22050,
sampling_rate=16000,
mel_fmin=0.0,
mel_fmax=8000.0):
"""
PyTorch layer which calculates mel spectrograms
Args:
filter_length (int): the number of fft components
hop_length (int): the window slide over the waveform,
with suggested value 12.5ms
win_length (int): the size of the window to be applied, suggested
value is 50ms
n_mel_channels (int): the number of mel bands/bins to be generated
sampling_rate (int): the sampling rate of the audio waveform, given
data of the same sr suggested value can be found via sox
mel_fmin (float): pls refer to librosa documentation
mel_fmax (flaot): pls refer to librosa documentation
"""
super(TacotronSTFT, self).__init__()
self.n_mel_channels = n_mel_channels
self.sampling_rate = sampling_rate
self.stft_fn = STFT(filter_length, hop_length, win_length)
mel_basis = librosa_mel_fn(
sampling_rate, filter_length, n_mel_channels, mel_fmin, mel_fmax)
mel_basis = torch.from_numpy(mel_basis).float()
self.register_buffer('mel_basis', mel_basis)
def __init__(self, filter_length=1024, hop_length=256, win_length=1024,
n_mel_channels=80, sampling_rate=22050, mel_fmin=0.0,
mel_fmax=8000.0):
super(TacotronSTFT, self).__init__()
self.n_mel_channels = n_mel_channels
self.sampling_rate = sampling_rate
self.stft_fn = STFT(filter_length, hop_length, win_length)
mel_basis = librosa_mel_fn(
sampling_rate, filter_length, n_mel_channels, mel_fmin, mel_fmax)
mel_basis = torch.from_numpy(mel_basis).float()
self.register_buffer('mel_basis', mel_basis)
def __init__(self,
filter_length=1024,
hop_length=256,
win_length=1024,
n_mel_channels=80,
sampling_rate=22050,
mel_fmin=50.0,
mel_fmax=7600.0):
super(TacotronSTFT, self).__init__()
self.n_mel_channels = n_mel_channels
self.sampling_rate = sampling_rate
self.stft_fn = STFT(filter_length, hop_length, win_length)
mel_basis = librosa.filters.mel(sampling_rate, filter_length,
n_mel_channels, mel_fmin, mel_fmax)
import numpy as np
inv_mel_basis = np.linalg.pinv(mel_basis)
mel_basis = torch.from_numpy(mel_basis).float()
inv_mel_basis = torch.from_numpy(inv_mel_basis).float()
self.register_buffer('mel_basis', mel_basis)
self.register_buffer('inv_mel_basis', inv_mel_basis)
class TacotronSTFT(torch.nn.Module):
def __init__(self,
filter_length=1024,
hop_length=256,
win_length=1024,
n_mel_channels=80,
sampling_rate=24000,
mel_fmin=0.0,
mel_fmax=12000.0):
super(TacotronSTFT, self).__init__()
self.n_mel_channels = n_mel_channels
self.sampling_rate = sampling_rate
self.stft_fn = STFT(filter_length, hop_length, win_length)
mel_basis = librosa_mel_fn(sampling_rate, filter_length,
n_mel_channels, mel_fmin, mel_fmax)
mel_basis = torch.from_numpy(mel_basis).float()
self.register_buffer('mel_basis', mel_basis)
def spectral_normalize(self, magnitudes):
output = dynamic_range_compression(magnitudes)
return output
def spectral_de_normalize(self, magnitudes):
output = dynamic_range_decompression(magnitudes)
return output
def mel_spectrogram(self, y):
"""Computes mel-spectrograms from a batch of waves
PARAMS
------
y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
RETURNS
-------
mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
"""
if ((torch.max(y.data) > 1) or (torch.min(y.data) < -1)):
y.data /= torch.max(torch.min(y.data).abs(), torch.max(y.data))
print(f"Normalized: {torch.min(y.data)} ... {torch.max(y.data)}")
assert (torch.min(y.data) >=
-1), f"Min value of audio tensor: {torch.min(y.data)} < -1"
assert (torch.max(y.data) <=
1), f"Max value of audio tensor: {torch.max(y.data)} > 1"
magnitudes, phases = self.stft_fn.transform(y)
magnitudes = magnitudes.data
mel_output = torch.matmul(self.mel_basis, magnitudes)
mel_output = self.spectral_normalize(mel_output)
return mel_output
class TacotronSTFT(torch.nn.Module):
def __init__(self,
filter_length=2048,
hop_length=275,
win_length=1100,
n_mel_channels=80,
sampling_rate=22050,
mel_fmin=125.0,
mel_fmax=7600.0):
super(TacotronSTFT, self).__init__()
self.n_mel_channels = n_mel_channels
self.sampling_rate = sampling_rate
self.stft_fn = STFT(filter_length, hop_length, win_length)
mel_basis = librosa.filters.mel(sampling_rate, filter_length,
n_mel_channels, mel_fmin, mel_fmax)
mel_basis = torch.from_numpy(mel_basis).float()
self.register_buffer('mel_basis', mel_basis)
def spectral_normalize(self, magnitudes):
output = dynamic_range_compression(magnitudes)
return output
def spectral_de_normalize(self, magnitudes):
output = dynamic_range_decompression(magnitudes)
return output
def mel_spectrogram(self, y):
"""Computes mel-spectrograms from a batch of waves
PARAMS
------
y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
RETURNS
-------
mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
"""
assert (torch.min(y.data) >= -1)
assert (torch.max(y.data) <= 1)
magnitudes, phases = self.stft_fn.transform(y)
magnitudes = magnitudes.data
mel_output = torch.matmul(self.mel_basis, magnitudes)
mel_output = self.spectral_normalize(mel_output)
return mel_output
class TacotronSTFT(torch.nn.Module):
def __init__(self,
filter_length=1024,
hop_length=256,
win_length=1024,
n_mel_channels=80,
sampling_rate=22050,
mel_fmin=50.0,
mel_fmax=7600.0):
super(TacotronSTFT, self).__init__()
self.n_mel_channels = n_mel_channels
self.sampling_rate = sampling_rate
self.stft_fn = STFT(filter_length, hop_length, win_length)
mel_basis = librosa.filters.mel(sampling_rate, filter_length,
n_mel_channels, mel_fmin, mel_fmax)
import numpy as np
inv_mel_basis = np.linalg.pinv(mel_basis)
mel_basis = torch.from_numpy(mel_basis).float()
inv_mel_basis = torch.from_numpy(inv_mel_basis).float()
self.register_buffer('mel_basis', mel_basis)
self.register_buffer('inv_mel_basis', inv_mel_basis)
def spectral_normalize(self, magnitudes):
return dynamic_range_compression(magnitudes)
def spectral_de_normalize(self, magnitudes):
return dynamic_range_decompression(magnitudes)
def mel_spectrogram(self, y):
"""Computes mel-spectrograms from a batch of waves
PARAMS
------
y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
RETURNS
-------
mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
"""
#assert(torch.min(y.data) >= -1)
#assert(torch.max(y.data) <= 1)
magnitudes, phases = self.stft_fn.transform(y)
magnitudes = magnitudes.data
mel_output = torch.matmul(self.mel_basis, magnitudes)
mel_output = self.spectral_normalize(mel_output)
return mel_output
def inv_mel_spectrogram(self, mel):
"""Computes mel-spectrograms from a batch of waves
PARAMS
------
y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
RETURNS
-------
mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
"""
mel = self.spectral_de_normalize(mel.float())
magnitudes = torch.matmul(self.inv_mel_basis, mel.data)
magnitudes = torch.max(magnitudes.clone().detach().fill_(1e-10),
magnitudes)
return magnitudes.data
class TacotronSTFT(torch.nn.Module):
def __init__(self,
filter_length=1024,
hop_length=256, win_length=1024,
n_mel_channels=80,
# sampling_rate=22050,
sampling_rate=16000,
mel_fmin=0.0,
mel_fmax=8000.0):
"""
PyTorch layer which calculates mel spectrograms
Args:
filter_length (int): the number of fft components
hop_length (int): the window slide over the waveform,
with suggested value 12.5ms
win_length (int): the size of the window to be applied, suggested
value is 50ms
n_mel_channels (int): the number of mel bands/bins to be generated
sampling_rate (int): the sampling rate of the audio waveform, given
data of the same sr suggested value can be found via sox
mel_fmin (float): pls refer to librosa documentation
mel_fmax (flaot): pls refer to librosa documentation
"""
super(TacotronSTFT, self).__init__()
self.n_mel_channels = n_mel_channels
self.sampling_rate = sampling_rate
self.stft_fn = STFT(filter_length, hop_length, win_length)
mel_basis = librosa_mel_fn(
sampling_rate, filter_length, n_mel_channels, mel_fmin, mel_fmax)
mel_basis = torch.from_numpy(mel_basis).float()
self.register_buffer('mel_basis', mel_basis)
# print(f"The mel channels are {self.n_mel_channels}")
# print(f"The sr is {self.sampling_rate}")
# print(f"The filter length is {filter_length}")
# print(f"The hop length is {hop_length}")
# print(f"The win length is {win_length}")
# print(f"The mel basis has size {self.mel_basis.size()}")
def spectral_normalize(self, magnitudes):
output = dynamic_range_compression(magnitudes)
return output
def spectral_de_normalize(self, magnitudes):
output = dynamic_range_decompression(magnitudes)
return output
def spectrogram(self, y):
"""
Computes spectrogram from batch of waves
Args:
y (torch.FloatTensor) (B, T) in range [-1, 1]
Returns:
magnitudes (torch.FloatTensor) (B, n_spec_channels, T)
"""
assert(torch.min(y.data) >= -1)
assert(torch.max(y.data) <= 1)
magnitudes, phases = self.stft_fn.transform(y)
return magnitudes
def mel_spectrogram(self, y):
"""Computes mel-spectrograms from a batch of waves
PARAMS
------
y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
RETURNS
-------
mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
"""
assert(torch.min(y.data) >= -1)
assert(torch.max(y.data) <= 1)
magnitudes, phases = self.stft_fn.transform(y)
magnitudes = magnitudes.data
mel_output = torch.matmul(self.mel_basis, magnitudes)
mel_output = self.spectral_normalize(mel_output)
return mel_output
