File size: 7,948 Bytes
9a70c5d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 |
'''
This is the main ArrayID feature building script
revised: April 04, 2021
'''
import glob
import os
import numpy as np
import matplotlib.pyplot as plt
from scipy.io.wavfile import read
from scipy.fftpack import fft, ifft, fftfreq
from scipy import signal
import random
from librosa.core import lpc
import librosa.feature
import csv
from sklearn.preprocessing import normalize
from direction_detection import *
##############################################
# HELPER FUNCTIONS
# converts hz to indices -> allows splicing of freq data
def hz_to_indices(freqs, lowcut, highcut):
i = 0
while freqs[i] < lowcut:
i += 1
low = i
while freqs[i] < highcut:
i += 1
return low, i
# compresses our feature vectors
# After extracting our features, they could be different lengths depending on
# the input signal, so we normalize each feature vector to be the same no matter
# the speaker
def get_row_compressor(old_dimension, new_dimension):
dim_compressor = np.zeros((new_dimension, old_dimension))
bin_size = float(old_dimension) / new_dimension
next_bin_break = bin_size
which_row = 0
which_column = 0
while which_row < dim_compressor.shape[0] and which_column < dim_compressor.shape[1]:
if round(next_bin_break - which_column, 10) >= 1:
dim_compressor[which_row, which_column] = 1
which_column += 1
elif next_bin_break == which_column:
which_row += 1
next_bin_break += bin_size
else:
partial_credit = next_bin_break - which_column
dim_compressor[which_row, which_column] = partial_credit
which_row += 1
dim_compressor[which_row, which_column] = 1 - partial_credit
which_column += 1
next_bin_break += bin_size
dim_compressor /= bin_size
return dim_compressor
# helper functions for above function
def get_column_compressor(old_dimension, new_dimension):
return get_row_compressor(old_dimension, new_dimension).transpose()
def compress_and_average(array, new_shape):
return np.mat(get_row_compressor(array.shape[0], new_shape[0])) * \
np.mat(array) * \
np.mat(get_column_compressor(array.shape[1], new_shape[1]))
##############################################
##############################################
# MAIN FEATURE EXTRACTION FUNCTIONS
def get_filelist(dir):
Filelist = []
for home, dirs, files in os.walk(dir):
for filename in files:
Filelist.append(os.path.join(home, filename))
return Filelist
def lpcc(data, n=15):
"""
f_LPC = lpcc(data, n): get the LPCC from the voice data
The order n is 15
"""
size_lpc = n # define the order of LPCC
a = lpc(data, order = size_lpc) # use the built-in function
a = -a
f_LPC = np.zeros(len(a))
f_LPC[0] = np.log(size_lpc)
for i in range(1, len(a)):
k = np.arange(1, i) # k from 1 to i-1
f_LPC[i] = a[i] + np.sum((1 - k/i) * a[k] * f_LPC[i - k])
return f_LPC[1:]
# returns long term fft
def get_ltfd(spec, m=20, start_index=1, end_index=86):
# only get the useful part
spec = spec[:, start_index: end_index, :(spec.shape[2] - spec.shape[2] % m)]
# merge the spec in the time line
channels = np.sum(spec, axis=2)
all_ffts = np.sum(channels, axis=0)
all_ffts /= np.max(all_ffts)
channels_ffts = np.asarray([channels[i, :] / np.max(channels[i, :]) for i in range(channels.shape[0])])
return all_ffts, channels_ffts
# returns long term fft
def get_ltfp(spec, m=20, start_index_fp=1, end_index_fp=86):
# only get the useful part
spec = spec[:, start_index_fp:end_index_fp, :(spec.shape[2] - spec.shape[2] % m)]
# split the data
splices = np.asarray(np.split(spec, m, axis=2))
# merge the data (wang ge hua)
mesh = np.zeros((splices.shape[0], splices.shape[1], splices.shape[2]))
for i in range(mesh.shape[0]):
for j in range(mesh.shape[1]):
for k in range(mesh.shape[2]):
mesh[i, j, k] = np.sum(splices[i, j, k, :])
# calculate the standard deviation
std_feature = np.zeros((mesh.shape[0], mesh.shape[2]))
for i in range(std_feature.shape[0]):
for j in range(std_feature.shape[1]):
std_feature[i, j] = np.std(mesh[i, :, j]) / np.mean(mesh[i, :, j])
# define the ltfp
LTFP = np.mean(std_feature, axis=0)
LTFP = LTFP / np.max(LTFP)
return LTFP
def feature_distribution(channel_fft):
num_feature = 5
f_dis = np.zeros(2 * num_feature)
co = np.zeros((num_feature, len(channel_fft)))
for num in range(len(channel_fft)):
a = channel_fft[num]
for i in range(1, len(a)):
a[i] = a[i-1] + a[i]
a = a / np.max(a)
dis_index = [0.1, 0.3, 0.5, 0.7, 0.9]
for i in range(len(dis_index)):
co[i, num] = find_value(a, dis_index[i])
co[:, num] /= len(a)
for i in range(num_feature):
f_dis[i] = np.mean(co[i, :])
f_dis[i + num_feature] = np.std(co[i, :])
return co, f_dis
def find_value(a, dis_index):
c = 0
for i in range(len(a) - 1):
if a[i] <= dis_index <= a[i + 1]:
c = i
break
return c
def mald_feature(rate, data):
n_fft = 4096
# detect the direction
if data.shape[1] == 4:
closestPair = getAngle_for_four(data, fs=rate)
elif data.shape[1] == 6:
closestPair = getAngle_for_six(data, fs=rate)
elif data.shape[1] == 8:
closestPair = getAngle_for_eight(data, fs=rate)
pairs = getDirection_Pair(closestPair, data.shape[1])
# low and high thresholds for field print features -> we want 1 - 10kHz range
lowcut_fp = 1
highcut_fp = 5000
if highcut_fp > rate / 2: # in case the sampling rate is very small
highcut_fp = rate / 2 - 100
highcut_fd = 1000
# input rate -> make sure to change this based on device.
# All of the devices are 44100 except for the AMLOGIC, which is 16kHz.
# If this rate is not changed acccordingly, the _ltfp and _ltfft features
# will be off
# just some helper splicing globals
freq = fftfreq(n_fft, 1. / rate) # data = logmmse(data, rate)
start_index, end_index = hz_to_indices(freq, lowcut_fp, highcut_fd)
start_index_fp, end_index_fp = hz_to_indices(freq, lowcut_fp, highcut_fp)
# empty feature vectors
_lpcc = []
# extract lfp and lpcc from each channel independently, then sum
for i in pairs:
a = np.asfortranarray(data[:, i]).astype(dtype=float)
_lpcc += list(lpcc(a))
# calculate the spectrogram
spec = [signal.stft(data[:, i], fs=rate, window='hann', nperseg=1024, noverlap=768, nfft=n_fft)[2] for i in range(data.shape[1])]
spec = np.asarray(spec) # convert list to numpy
# obtain the absolute value
spec = np.abs(spec)
# get the ltfp feature
# get ltfp features and compress to a 50 feature vectoc
_ltfd, channel_fft = get_ltfd(spec=spec, start_index=start_index, end_index=end_index)
_ltfd = list(compress_and_average(_ltfd.reshape(len(_ltfd), 1), (20, 1)).flat)
co, _fdis = feature_distribution(channel_fft)
# get ltfp features and compress to a 50 feature vector
_ltfp = get_ltfp(spec=spec, start_index_fp=start_index_fp, end_index_fp=end_index_fp)
_ltfp = list(compress_and_average(_ltfp.reshape(len(_ltfp), 1), (20, 1)).flat)
# out is final feature vector, each data point formed as a tuple : (X, y), where X is the feature vector and y is the label
# X_y is just compiled l ist of all the tuples
feature = np.concatenate((_lpcc, _ltfd, _fdis, _ltfp))
return feature
|