def preprocess_data(self, x):
# if IS_CUT_AUDIO:
# x = [sample[0:MAX_AUDIO_DURATION*AUDIO_SAMPLE_RATE] for sample in x]
# extract mfcc
x = extract_mfcc_parallel(x, n_mfcc=96)
if self.max_length is None:
self.max_length = get_max_length(x)
self.max_length = min(MAX_FRAME_NUM, self.max_length)
x = pad_seq(x, pad_len=self.max_length)
return x
def preprocess_data(self, x):
# extract mfcc
x = extract_mfcc(x)
if self.max_length is None:
self.max_length = get_max_length(x)
x = pad_seq(x, self.max_length)
# feature scale
if self.mean is None or self.std is None:
self.mean = np.mean(x)
self.std = np.std(x)
x = (x - self.mean) / self.std
# calculate mean of mfcc
x = np.mean(x, axis=-1)
x = x[:, :, np.newaxis]
return x
def preprocess_data(self, x):
if IS_CUT_AUDIO:
x = [
sample[0:MAX_AUDIO_DURATION * AUDIO_SAMPLE_RATE]
for sample in x
]
x_mel = extract_melspectrogram_parallel(x,
n_mels=128,
use_power_db=True)
# x_mel = extract_mfcc_parallel(x, n_mfcc=96)
if self.max_length is None:
self.max_length = get_max_length(x_mel)
self.max_length = min(MAX_FRAME_NUM, self.max_length)
x_mel = pad_seq(x_mel, pad_len=self.max_length)
x_mel = x_mel[:, :, :, np.newaxis]
return x_mel
def preprocess_data(self, x):
# mel-spectrogram parameters
SR = 16000
N_FFT = 512
N_MELS = 96
HOP_LEN = 256
DURA = 21.84 # to make it 1366 frame.
if IS_CUT_AUDIO:
x = [sample[0:MAX_AUDIO_DURATION * AUDIO_SAMPLE_RATE]
for sample in x]
# x_mel = extract_melspectrogram_parallel(x, n_mels=128, use_power_db=True)
x_mfcc = extract_mfcc_parallel(x, n_mfcc=96)
if self.max_length is None:
self.max_length = get_max_length(x_mfcc)
self.max_length = min(MAX_FRAME_NUM, self.max_length)
x_mfcc = pad_seq(x_mfcc, pad_len=self.max_length)
x_mfcc = x_mfcc[:, :, :, np.newaxis]
return x_mfcc
def preprocess_data(self, x):
if IS_CUT_AUDIO:
x = [
sample[0:MAX_AUDIO_DURATION * AUDIO_SAMPLE_RATE]
for sample in x
]
# extract mfcc
x = extract_mfcc_parallel(x, n_mfcc=96)
if self.max_length is None:
self.max_length = get_max_length(x)
x = pad_seq(x, self.max_length)
# if self.scaler is None:
# self.scaler = []
# for i in range(x.shape[2]):
# self.scaler.append(StandardScaler().fit(x[:, :, i]))
# for i in range(x.shape[2]):
# x[:, :, i] = self.scaler[i].transform(x[:, :, i])
# feature scale
# if self.mean is None or self.std is None:
# self.mean = np.mean(x)
# self.std = np.std(x)
# x = (x - self.mean) / self.std
# s0, s1, s2 = x.shape[0], x.shape[1], x.shape[2]
# x = x.reshape(s0 * s1, s2)
# if not self.scaler:
# self.scaler = MinMaxScaler().fit(x)
# x = self.scaler.transform(x)
# x = x.reshape(s0, s1, s2)
# 4 dimension?
# (120, 437, 24) to (120, 437, 24, 1)
# 120 is the number of instance
# 437 is the max length
# 24 frame in mfcc
# log(f"max {np.max(x)} min {np.min(x)} mean {np.mean(x)}")
x = x[:, :, :, np.newaxis]
return x
def preprocess_data(self, x):
if IS_CUT_AUDIO:
x = [
sample[0:MAX_AUDIO_DURATION * AUDIO_SAMPLE_RATE]
for sample in x
]
# extract mfcc
x_mfcc = extract_mfcc_parallel(x, n_mfcc=64)
x_mel = extract_melspectrogram_parallel(x,
n_mels=64,
use_power_db=True)
if self.max_length is None:
self.max_length = get_max_length(x_mfcc)
self.max_length = min(MAX_FRAME_NUM, self.max_length)
x_mfcc = pad_seq(x_mfcc, self.max_length)
x_mel = pad_seq(x_mel, self.max_length)
x_feas = np.concatenate([x_mfcc, x_mel], axis=-1)
x_feas = x_feas[:, :, :, np.newaxis]
# x_mel = pad_seq(x_mel, self.max_length)
# x_mel = x_mel[:, :, :, np.newaxis]
return x_feas