def execute(silences,
inputFolder='./raw/',
outputFolder='./train/',
segment=500,
verbose=True,
defaultSilenceLen=defaultLen):
import random as ran
silences_list_len = len(silences) # number of silences sample
for file in listFile(os.walk(inputFolder), ['*.wav']):
org = rd.importSignal(inputFolder + file)
wav, __len = trim(org, outOrgLen=True)
for i in range(
1
): # range (max ((defaultSilenceLen - __len) // segment-2, 1)):
idx = ran.randrange(0, silences_list_len)
choice = silences[idx]
seg = mix(wav, choice, i * segment)
sname = outputFolder + 'mix_%d_%d_%s' % (idx, i, file)
rd.save(seg, sname)
if verbose:
print("Saved file:%-40s with sample size:%5d" %
(sname, len(seg)))
def main():
# print (X_train)
if os.path.isfile('model.keras'):
model = load_lobe()
else:
model = network_1()
# for train purpose only
# follow this standard
X_train, Y_train = loadData()
X_train = X_train.reshape(X_train.shape[0], X_train.shape[1], 1)
#
model.fit(X_train, Y_train, batch_size=32, nb_epoch=50)
save_lobe(model)
for file in listFile(os.walk('./test/'), ['*nhu*.wav']):
mfcc = loadMfccFromFile('./test/' + file)
mfcc = mfcc.reshape(1, mfcc.shape[0], 1)
pred = model.predict(mfcc)
print('Predict for %-40s:' % file, pred, pred[0][0] > 0.5)
pass
def loadSilences(dir='./raw/silent/'):
silences = []
for silent in listFile(os.walk(dir), ['*.wav']):
silence = rd.importSignal('%s%s' % (dir, silent))[0:defaultLen]
silences.append(silence)
# print (len (silence))
return silences
def loadFolder(nhu, label):
import json
if os.path.isfile(nhu + 'cache'):
a = np.loadtxt(nhu + 'cache')
return np.array(a), np.array([label] * len(a))
else:
X_train = []
Y_train = []
flist = listFile(os.walk(nhu), ['*.wav'])
i = 0
for file in flist:
i += 1
print('%4d/%4d. Loading %-40s for label %5d' %
(i, len(flist), nhu + file, label))
mfcc = loadMfccFromFile(nhu + file)
X_train.append(mfcc)
Y_train.append(label)
# print (X_train)
np.savetxt(nhu + 'cache', X_train)
return np.array(X_train), np.array(Y_train)
def main():
# ask ()
# return;
# os.environ["CUDA_VISIBLE_DEVICES"]="-1"
# from play import listFile
# print (listFile (pattern=['nhu*']))
# return
from play import listFile
# print (listFile (pattern=['khong*']))
lobe = None
if not os.path.isfile('model.keras'):
lobe = voice_lobe_5th()
print(lobe.summary())
# return;
yes, no = collect()
print('Importing Data set ...')
X_train, Y_train = importData(yes, no)
# Y_train = Y_train.reshape(Y_train.shape[0], 1, 1)
# print (Y_train.shape)
# Y_train = Y_train.reshape (1, Y_train.shape[0], 1)
# X_train = X_train.reshape (len(X_train), defaultLen // particle, particle)
# Y_train = Y_train.reshape (len(Y_train), 1)
# print (X_train)
# return;
# print (X_train)
lobe.fit(X_train, Y_train, batch_size=len(X_train), nb_epoch=300)
save_lobe(lobe)
else:
lobe = load_lobe()
# return;
# from ann_visualizer.visualize import ann_viz;
# ann_viz(lobe, title="My second neural network")
def check(i):
from time import time
wav = rd.importSignal(i)
mfcc = transRaw(wav, trim=True) # have to trim a word
#from time import sleep
mfcc = mfcc.reshape(1, mfcc.shape[0], mfcc.shape[1])
# mfcc = mfcc.reshape (1, defaultLen // particle, particle)
mark = time()
pred = lobe.predict(mfcc)
print('%-40s' % str(pred),
'for %10s estimate time in second: %5.4f' % (i, time() - mark),
end=' ')
chosen = 0
mx = 0
for i in range(len(pred[0])):
if mx < pred[0][i]:
chosen = i
mx = pred[0][i]
print('label:', chosen)
return (pred[0][0] >= 0.5)
# for i in listFile (pattern=['nhu*']):
# check (i)
# return;
cnt = 0
_len = 0
for i in listFile(pattern=['nhu*', 'khong*', 'toi*']):
if check(i):
cnt += 1
_len += 1