def main():
theta1 = loadmat('ml.mat')['theta1']
theta2 = loadmat('ml.mat')['theta2']
"""Xtest = [];
ytest = [];
nspeakers = theta2.shape[0];
folders = os.listdir("wav")
for i in range(5):
folder = folders[i];
print(folder)
files = [f for f in glob.glob("wav/"+folder + "/" + "**/*.wav", recursive=True)]
sztraining = int(len(files)*0.6);
for fid in range(sztraining, len(files)):
sample_rate, signal = wav.read(files[fid])
signal = signal[0:int(2 * sample_rate)]
mfcc = MFCC.main(signal, sample_rate)
Xtest.append(mfcc)
ytest.append(i)
ytest = np.array(ytest)
Xtest = np.array(Xtest)
pred = [];
for i in range(len(Xtest)):
pred.append(ml.predictWAV(theta1, theta2, Xtest[i])[0])
print(np.mean(pred == ytest.flatten()) * 100)"""
signal = []
sample_rate = 16000
#th = threading.Thread(target=audio.plot_audio, args=(1,));
#th.start()
while True:
cmd = input("Digite um comando")
print("CMDZAO = " + str(cmd))
if cmd == "record":
seconds = 7
print("recording...")
signal = sd.rec(int(seconds * sample_rate),
samplerate=sample_rate,
channels=1)
sd.wait()
elif cmd == "who":
if not len(signal):
print("no signal")
continue
sd.play(signal, sample_rate)
signal = signal[0:int(2 * sample_rate)]
mfcc = MFCC.main(signal, sample_rate)
mlres = ml.predictWAV(theta1, theta2, mfcc)
print("user id: {}".format(mlres[0]))
elif cmd == "exit":
break
else:
print("not found.")
return 0
def getPerson(self):
if self.file[0] == '':
print('empty file!')
return
sample_rate, signal = read(self.file[0])
theta1 = loadmat('ml.mat')['theta1']
theta2 = loadmat('ml.mat')['theta2']
mfcc = MFCC.main(signal, sample_rate)
mlres = ml.predictWAV(theta1, theta2, mfcc)
self.whoAmIWindow = WhoAmIWindowClass(mlres)
self.whoAmIWindow.setWindowTitle('Result')
self.whoAmIWindow.show()
def main():
folders = os.listdir("wav")
X = []
y = []
Xtest = []
ytest = []
nspeakers = 5
#feature extraction
for i in range(nspeakers):
folder = folders[i]
files = [
f for f in glob.glob("wav/" + folder + "/" + "**/*.wav",
recursive=True)
]
sztraining = int(len(files) * 0.6)
for fid in range(sztraining):
sample_rate, signal = wav.read(files[fid])
mfcc = MFCC.main(signal, sample_rate)
for j in range(len(mfcc)):
X.append([])
for k in range(len(mfcc[j])):
X[-1].append(mfcc[j][k])
y.append(i)
for fid in range(sztraining, len(files)):
sample_rate, signal = wav.read(files[fid])
mfcc = MFCC.main(signal, sample_rate)
for j in range(len(mfcc)):
Xtest.append([])
for k in range(len(mfcc[j])):
Xtest[-1].append(mfcc[j][k])
ytest.append(i)
y = np.array(y)
X = np.array(X)
ytest = np.array(ytest)
Xtest = np.array(Xtest)
input_layer_size = 390
hidden_layer_size = 200
num_labels = nspeakers
lmbda = 1
initial_theta1 = ml.randInitializeWeights(input_layer_size,
hidden_layer_size)
initial_theta2 = ml.randInitializeWeights(hidden_layer_size, num_labels)
nn_initial_params = np.hstack(
(initial_theta1.ravel(order='F'), initial_theta2.ravel(order='F')))
print(
ml.nnCostFunc(nn_initial_params, input_layer_size, hidden_layer_size,
num_labels, X, y, lmbda))
theta_opt = opt.fmin_cg(maxiter=50,
f=ml.nnCostFunc,
x0=nn_initial_params,
fprime=ml.nnGrad,
args=(input_layer_size, hidden_layer_size,
num_labels, X, y.flatten(), lmbda))
theta1_opt = np.reshape(
theta_opt[:hidden_layer_size * (input_layer_size + 1)],
(hidden_layer_size, input_layer_size + 1), 'F')
theta2_opt = np.reshape(
theta_opt[hidden_layer_size * (input_layer_size + 1):],
(num_labels, hidden_layer_size + 1), 'F')
pred = ml.predict(theta1_opt, theta2_opt, Xtest, ytest)
print(np.mean(pred == ytest.flatten()) * 100)
savemat('ml.mat', {
'theta1': theta1_opt,
'theta2': theta2_opt
})
