def frame_combine(frame, file_path, start, end):
fbank = fbank_reader.HTKFeat_read(file_path).getall()
if end - start + 1 < 41:
if frame - start <= 30 and end - frame <= 10:
frame_to_combine = []
front_rest = 30 - (frame - start)
back_rest = 10 - (end - frame)
for i in range(front_rest):
frame_to_combine.append(fbank[start].tolist())
for i in range(start, end + 1):
frame_to_combine.append(fbank[i].tolist())
for i in range(back_rest):
frame_to_combine.append(fbank[end].tolist())
elif end - frame >= 10:
frame_to_combine = []
front_rest = 30 - (frame - start)
for i in range(front_rest):
frame_to_combine.append(fbank[start].tolist())
for i in range(start, frame + 11):
frame_to_combine.append(fbank[i].tolist())
else:
frame_to_combine = []
back_rest = 10 - (end - frame)
for i in range(frame - 30, end + 1):
frame_to_combine.append(fbank[i].tolist())
for i in range(back_rest):
frame_to_combine.append(fbank[end].tolist())
combined = np.array(frame_to_combine).reshape(-1)
else:
if frame - start >= 30 and end - frame >= 10:
frame_to_combine = fbank[frame - 30:frame + 11]
combined = frame_to_combine.reshape(-1)
elif frame - start < 30:
frame_to_combine = fbank[start:start + 41]
combined = frame_to_combine.reshape(-1)
else:
frame_to_combine = fbank[end - 40:end + 1]
combined = frame_to_combine.reshape(-1)
return combined.tolist()
def read_train_set(self):
self.example = []
self.labels = []
self.num_examples = 0
step_length = 10
start = self.fbank_position % len(self.examples_list)
end = (self.fbank_position + step_length) % len(self.examples_list)
if start < end:
fbank_list = self.examples_list[start:end]
self.fbank_position += step_length
else:
fbank_list = self.examples_list[start:len(self.examples_list)]
self.fbank_position = 0
index = np.arange(len(self.examples_list))
np.random.shuffle(index)
self.examples_list = np.array(self.examples_list)[index]
for example in fbank_list:
if example == '':
continue
file_path = example.split(" ")[1]
if example.split('/').count("positive") > 0:
start = self.position_data.find(getname(example))
end = self.position_data.find("positive", start + 1)
if end != -1:
position_str = self.position_data[start + 15:end - 1]
else:
position_str = self.position_data[start + 15:end]
# start and end position of "hello" & start and end position of "xiao gua"
keyword_position = position_str.split(" ")
file_path = example.split(" ")[1]
keyword_frame_position = []
for i in range(4):
fbank = fbank_reader.HTKFeat_read(file_path).getall()
length = fbank.shape[0]
frame_position = (int(keyword_position[i]) - 240) // 160
if frame_position >= length:
frame_position = length - 1
keyword_frame_position.append(frame_position)
#print (example)
for frame in range(keyword_frame_position[0],
keyword_frame_position[1] + 1):
self.example.append(
frame_combine(frame, file_path,
keyword_frame_position[0],
keyword_frame_position[1]))
self.labels.append('0')
self.num_examples += 1
for frame in range(keyword_frame_position[2],
keyword_frame_position[3] + 1):
self.example.append(
frame_combine(frame, file_path,
keyword_frame_position[2],
keyword_frame_position[3]))
self.labels.append('1')
self.num_examples += 1
else:
file_path = example.split(" ")[1]
fbank = fbank_reader.HTKFeat_read(file_path).getall()
frame_number = fbank.shape[0]
#print (example)
for frame in range(frame_number):
self.example.append(
frame_combine(frame, file_path, 0, frame_number - 1))
self.labels.append('2')
self.num_examples += 1
def get_next_batch(self, batch_size):
example = []
label = []
start = self.position % self.length
if start == 0:
index = np.array(range(self.length))
np.random.shuffle(index)
self.file_list = np.array(self.file_list)[index]
end = (self.position + batch_size) % self.length
if start < end:
fbank_list = self.file_list[start:end]
self.position += batch_size
else:
fbank_list = self.file_list[start:self.length]
self.position = 0
for fileinfo in fbank_list:
l = []
ll = []
file_path = fileinfo.split(" ")[1]
if fileinfo.split("/").count("negative") > 0:
fbank = fbank_reader.HTKFeat_read(file_path).getall()
for i in range(fbank.shape[0]):
l.append(frame_combine(i, file_path, 0,
fbank.shape[0] - 1))
ll.append('2')
else:
fbank = fbank_reader.HTKFeat_read(file_path).getall()
first = self.position_list.find(getname(fileinfo))
second = self.position_list.find("positive", first + 1)
if second != -1:
second -= 1
position_str = self.position_list[first + 15:second]
keyword_position = position_str.split(" ")
keyword_frame_position = []
for i in range(4):
frame_position = (int(keyword_position[i]) - 240) // 160
if frame_position >= fbank.shape[0]:
frame_position = fbank.shape[0] - 1
keyword_frame_position.append(frame_position)
st = 0
en = keyword_frame_position[0]
for frame in range(st, en):
l.append(frame_combine(frame, file_path, st, en - 1))
ll.append('2')
st = en
en = keyword_frame_position[1] + 1
for frame in range(st, en):
l.append(frame_combine(frame, file_path, st, en - 1))
ll.append('0')
st = en
en = keyword_frame_position[2]
for frame in range(st, en):
l.append(frame_combine(frame, file_path, st, en - 1))
ll.append('2')
st = en
en = keyword_frame_position[3] + 1
for frame in range(st, en):
l.append(frame_combine(frame, file_path, st, en - 1))
ll.append('1')
st = en
en = fbank.shape[0]
for frame in range(st, en):
l.append(frame_combine(frame, file_path, st, en - 1))
ll.append('2')
# print("lshape",len(l))
example.append(np.array(l.copy()))
#print("example",len(example[0]))
label.append(np.array(ll.copy()))
res1 = np.array(example)
res2 = np.array(label)
print(res1.shape, res2.shape)
print(res1[1].shape)
return res1, res2
def read_data_sets():
f = open(
"/home/disk2/internship_anytime/aslp_hotword_data/aslp_wake_up_word_data/positiveKeywordPosition.txt",
"r")
position_data = f.read()
f.close()
f = open(
"/home/disk2/internship_anytime/aslp_hotword_data/aslp_wake_up_word_data/train.scp",
"r")
temp = f.read()
train_list = temp.split('\n')
f.close()
f = open(
"/home/disk2/internship_anytime/aslp_hotword_data/aslp_wake_up_word_data/test.scp",
"r")
temp = f.read()
test_list = temp.split('\n')
f.close()
test_examples = []
test_labels = []
test_length = []
test_num = 0
for example in test_list:
if example == '':
continue
if example.split('/').count("positive") < 1:
continue
start = position_data.find(getname(example))
end = position_data.find("positive", start + 1)
if end != -1:
position_str = position_data[start + 15:end - 1]
else:
position_str = position_data[start + 15:end]
# start and end position of "hello" & start and end position of "xiao gua"
keyword_position = position_str.split(" ")
file_path = example.split(" ")[1]
keyword_frame_position = []
for i in range(4):
fbank = fbank_reader.HTKFeat_read(file_path).getall()
length = fbank.shape[0]
frame_position = int(keyword_position[i]) // 160
if frame_position >= length:
frame_position = length - 1
keyword_frame_position.append(frame_position)
test_length.append(keyword_frame_position[1] -
keyword_frame_position[0] + 1 +
keyword_frame_position[3] -
keyword_frame_position[2] + 1)
# print( example)
for frame in range(keyword_frame_position[0],
keyword_frame_position[1] + 1):
test_examples.append(
frame_combine(frame, file_path, keyword_frame_position[0],
keyword_frame_position[1]))
test_labels.append('0')
test_num += 1
for frame in range(keyword_frame_position[2],
keyword_frame_position[3] + 1):
test_examples.append(
frame_combine(frame, file_path, keyword_frame_position[2],
keyword_frame_position[3]))
test_labels.append('1')
test_num += 1
# print(np.array(test_examples).shape)
for example in test_list:
if example == '':
continue
if example.split('/').count("negative") < 1:
continue
file_path = example.split(" ")[1]
fbank = fbank_reader.HTKFeat_read(file_path).getall()
frame_number = fbank.shape[0]
test_length.append(frame_number)
# print (example)
for frame in range(frame_number):
test_examples.append(
frame_combine(frame, file_path, 0, frame_number - 1))
test_labels.append('2')
test_num += 1
fbank_end_frame = []
for i in range(len(test_length)):
fbank_end_frame.append(sum(test_length[0:i + 1]))
index = np.arange(len(train_list))
np.random.shuffle(index)
train_list = np.array(train_list)[index]
train = TrainSet(train_list, position_data)
test = TestSet(test_examples, test_labels, test_num, fbank_end_frame)
return train, test