keunwoochoi

def istft(stft_matrix, hop_length=None, win_length=None, window='hann',
          center=True, normalized=False, onesided=True, length=None):
    """stft_matrix = (batch, freq, time, complex) 
    
    All based on librosa
        - http://librosa.github.io/librosa/_modules/librosa/core/spectrum.html#istft
    What's missing?
        - normalize by sum of squared window --> do we need it here?
        Actually the result is ok by simply dividing y by 2. 
    """

keunwoochoi

import torch
import torch.nn.functional as F
EPS = 1e-7

def _enhance_either_hpss(x_padded, out, kernel_size, power, which, offset):
    """x_padded: one that median filtering can be directly applied
    kernel_size = int
    dim: either 2 (freq-axis) or 3 (time-axis)

    which: str, either "harm" or "perc"

keunwoochoi

import torch



def STFT(src, n_fft, hop_length=None, window=None, **kwargs):
    """A wrapper for torch.stft with some preset parameters.
    Returned value keeps both real and imageinary parts.
    For STFT magnitude, see spectrogram

    """

keunwoochoi

class Net(nn.Module):
    def __init__(self, src):
        super(net, self).__init__() 
        self.src = src
        self.a = a
        self.conv1 = nn.Conv1d(1, 2, 3, 4,..)
    
    def _fwd_1(self, x):
        x = self.conv1(x)
        x = x + self.a

keunwoochoi

cqt_filter_fft = librosa.constantq.__cqt_filter_fft

class PseudoCqt():
    """A class to compute pseudo-CQT with Pytorch.
    Written by Keunwoo Choi
    API (+implementations) follows librosa (https://librosa.github.io/librosa/generated/librosa.core.pseudo_cqt.html)
    
    Usage:
        src, _ = librosa.load(filename)
        src_tensor = torch.tensor(src)

keunwoochoi

"""
Code converted by Keunwoo from https://ccrma.stanford.edu/~jos/fp/Formant_Filtering_Example.html (JOS)
"""

import numpy as np
import scipy.signal as signal
import librosa

def gen_speech(F, sig=None, filename='speech'):
    nsecs = len(F)

keunwoochoi

import matplotlib
%matplotlib inline
import matplotlib.pyplot as plt
plt.style.use('ggplot')
import numpy as np
import librosa
import keras
from keras import backend as K
print(keras.__version__)
print('Keras backend: ', keras.backend._BACKEND, ' and image dim ordering: ', keras.backend.image_dim_ordering())

keunwoochoi

import keras.backend as K

# Load the weights
# If there are more than 1 BN, you might wanna have a proper layer
moving_mean = f['batchnormalization_1_running_mean']
moving_std = f['batchnormalization_1_running_std']
beta = f['batchnormalization_1_beta']
gamma = f['batchnormalization_2_beta']

# BE CAREFUL! Keras 1 stores std

keunwoochoi

import keras
from keras import backend as K
from keras.layers.convolutional import Conv2D, Conv2DTranspose
from keras.layers import Input, Dense, Activation
from keras.layers import concatenate  # functional interface
from keras.models import Model
from keras.layers.advanced_activations import LeakyReLU

N_INPUT = 512

keunwoochoi

links = '''https://ismir2017.smcnus.org/wp-content/uploads/2017/10/126_Paper.pdf
https://ismir2017.smcnus.org/wp-content/uploads/2017/10/79_Paper.pdf
https://ismir2017.smcnus.org/wp-content/uploads/2017/10/89_Paper.pdf
https://ismir2017.smcnus.org/wp-content/uploads/2017/10/119_Paper.pdf
https://ismir2017.smcnus.org/wp-content/uploads/2017/10/51_Paper.pdf
https://ismir2017.smcnus.org/wp-content/uploads/2017/10/85_Paper.pdf
https://ismir2017.smcnus.org/wp-content/uploads/2017/10/164_Paper.pdf
https://ismir2017.smcnus.org/wp-content/uploads/2017/10/220_Paper.pdf
https://ismir2017.smcnus.org/wp-content/uploads/2017/10/172_Paper.pdf
https://ismir2017.smcnus.org/wp-content/uploads/2017/10/180_Paper.pdf