def ssc(signal,
samplerate=16000,
win_length=0.025,
win_step=0.01,
filters_num=26,
NFFT=512,
low_freq=0,
high_freq=None,
pre_emphasis_coeff=0.97):
'''
待补充
'''
high_freq = high_freq or samplerate / 2
signal = pre_emphasis(signal, pre_emphasis_coeff)
frames = audio2frame(signal, win_length * samplerate,
win_step * samplerate)
spec_power = spectrum_power(frames, NFFT)
spec_power = numpy.where(spec_power == 0,
numpy.finfo(float).eps, spec_power) #能量谱
fb = get_filter_banks(filters_num, NFFT, samplerate, low_freq, high_freq)
feat = numpy.dot(spec_power, fb.T) #计算能量
R = numpy.tile(
numpy.linspace(1, samplerate / 2, numpy.size(spec_power, 1)),
(numpy.size(spec_power, 0), 1))
return numpy.dot(spec_power * R, fb.T) / feat
def fbank(signal,
samplerate=16000,
win_length=0.025,
win_step=0.01,
filters_num=26,
NFFT=512,
low_freq=0,
high_freq=None,
pre_emphasis_coeff=0.97):
'''计算音频信号的MFCC
参数说明：
samplerate:采样频率
win_length:窗长度
win_step:窗间隔
filters_num:梅尔滤波器个数
NFFT:FFT大小
low_freq:最低频率
high_freq:最高频率
pre_emphasis_coeff:预加重系数
'''
high_freq = high_freq or samplerate / 2 #计算音频样本的最大频率
signal = pre_emphasis(signal, pre_emphasis_coeff) #对原始信号进行预加重处理
frames = audio2frame(signal, win_length * samplerate,
win_step * samplerate) #得到帧数组
spec_power = spectrum_power(frames, NFFT) #得到每一帧FFT以后的能量谱
energy = numpy.sum(spec_power, 1) #对每一帧的能量谱进行求和
energy = numpy.where(energy == 0,
numpy.finfo(float).eps,
energy) #对能量为0的地方调整为eps，这样便于进行对数处理
fb = get_filter_banks(filters_num, NFFT, samplerate, low_freq,
high_freq) #获得每一个滤波器的频率宽度
feat = numpy.dot(spec_power, fb.T) #对滤波器和能量谱进行点乘
feat = numpy.where(feat == 0, numpy.finfo(float).eps, feat) #同样不能出现0
return feat, energy
def getDataSet(wav_files,
label,
img_rows=864,
FrameLen=256,
FrameInc=64,
NFFT=800,
winfunc=lambda x: np.ones((x, ))):
X = []
Y = []
for i, singleWav in enumerate(wav_files):
(rate, sig) = wav.read(singleWav)
frames = audio2frame(sig, FrameLen, FrameInc,
winfunc) #blackman window alpha=0.16
logPowSpec = log_spectrum_power(frames, NFFT, norm=1)
#logPowSpec = spectrum_power(frames,NFFT)
#logPowSpec = log_fbank(sig,win_length=0.025,win_step=0.01,filters_num=39,NFFT)
#logPowSpec = calcMFCC_delta_delta(sig,win_length=0.025,win_step=0.01,cep_num=20,filters_num=26,NFFT)
dim1, dim2 = logPowSpec.shape
if dim1 >= img_rows:
divNum = dim1 / img_rows
logPowSpec = logPowSpec[0:divNum * img_rows, :]
logPowSpec = logPowSpec.reshape(divNum, img_rows, -1)
X.extend(logPowSpec)
Y.extend([label[i]] * divNum)
else:
divNum = img_rows / dim1
mod = img_rows % dim1
logPowSpec = np.array([logPowSpec] * divNum).reshape(
(divNum * dim1, dim2))
logPowSpec = np.concatenate((logPowSpec, logPowSpec[0:mod, :]))
X.append(logPowSpec)
Y.append(label[i])
return X, Y
def FFTcoefficient(sig,
samplerate=16000,
win_length=0.025,
win_step=0.01,
pre_emphasis_coeff=0.97,
NFFT=512):
'''
计算初始IDCT系数
:param sig:
:param samplerate:
:param win_length:
:param win_step:
:param pre_emphasis_coeff:
:return:
'''
#预处理
signal = pre_emphasis(sig, pre_emphasis_coeff)
#分帧
frames = audio2frame(signal, win_length * samplerate,
win_step * samplerate) # 得到帧数组
#加窗
frames *= np.hamming(int(round(win_length * samplerate))) # 加窗
#FFT
fftfeat = spectrum_power(frames, NFFT) # 进行快速傅里叶变换 得到幅值系数
feat = np.where(fftfeat == 0, np.finfo(float).eps, fftfeat)
#TODO 滤波
feat = np.log(fftfeat)
feat = dct(feat, type=2, axis=1, norm='ortho')
return feat
def f_spectrum_power(self, wavefile):
assert os.path.isfile(wavefile)
(rate, sig) = self.__readwav(wavefile)
high_freq = rate / 2 # 计算音频样本的最大频率
signal = pre_emphasis(sig, self.p_pre_emphasis_coeff) # 对原始信号进行预加重处理
frames = audio2frame(signal, self.p_win_len * rate,
self.p_win_step * rate) # 得到帧数组
frames *= np.hamming(int(round(self.p_win_len * rate))) # 加窗
spec_power = spectrum_power(frames, self.p_nfft) # 得到每一帧FFT以后的能量谱
return spec_power
def fbank(signal,
samplerate=16000,
win_length=0.025,
win_step=0.01,
filters_num=26,
NFFT=512,
low_freq=0,
high_freq=None,
pre_emphasis_coeff=0.97):
"""Perform pre-emphasis -> framing -> get magnitude -> FFT -> Mel Filtering.
Args:
signal: 1-D numpy array.
samplerate: Sampling rate. Defaulted to 16KHz.
win_length: Window length. Defaulted to 0.025, which is 25ms/frame.
win_step: Interval between the start points of adjacent frames.
Defaulted to 0.01, which is 10ms.
cep_num: Numbers of cepstral coefficients. Defaulted to 13.
filters_num: Numbers of filters. Defaulted to 26.
NFFT: Size of FFT. Defaulted to 512.
low_freq: Lowest frequency.
high_freq: Highest frequency.
pre_emphasis_coeff: Coefficient for pre-emphasis. Pre-emphasis increase
the energy of signal at higher frequency. Defaulted to 0.97.
Returns:
feat: Features.
energy: Energy.
"""
# Calculate the highest frequency.
high_freq = high_freq or samplerate / 2
# Pre-emphasis
signal = pre_emphasis(signal, pre_emphasis_coeff)
# rames: 2-D numpy array with shape (frame_num, frame_length)
frames = audio2frame(signal, win_length * samplerate,
win_step * samplerate)
# Caculate energy and modify all zeros to eps.
spec_power = spectrum_power(frames, NFFT)
energy = numpy.sum(spec_power, 1)
energy = numpy.where(energy == 0, numpy.finfo(float).eps, energy)
# Get Mel filter banks.
fb = get_filter_banks(filters_num, NFFT, samplerate, low_freq, high_freq)
# Get MFCC and modify all zeros to eps.
feat = numpy.dot(spec_power, fb.T)
feat = numpy.where(feat == 0, numpy.finfo(float).eps, feat)
return feat, energy
def fbank(signal,samplerate=16000,win_length=0.025,win_step=0.01,filters_num=26,NFFT=512,low_freq=0,high_freq=None,pre_emphasis_coeff=0.97):
'''计算音频信号的MFCC
参数说明：
samplerate:采样频率
win_length:窗长度
win_step:窗间隔
filters_num:梅尔滤波器个数
NFFT:FFT大小
low_freq:最低频率
high_freq:最高频率
pre_emphasis_coeff:预加重系数
'''
high_freq=high_freq or samplerate/2 #计算音频样本的最大频率
signal=pre_emphasis(signal,pre_emphasis_coeff) #对原始信号进行预加重处理
frames=audio2frame(signal,win_length*samplerate,win_step*samplerate) #得到帧数组
spec_power=spectrum_power(frames,NFFT) #得到每一帧FFT以后的能量谱
energy=numpy.sum(spec_power,1) #对每一帧的能量谱进行求和
energy=numpy.where(energy==0,numpy.finfo(float).eps,energy) #对能量为0的地方调整为eps，这样便于进行对数处理
fb=get_filter_banks(filters_num,NFFT,samplerate,low_freq,high_freq) #获得每一个滤波器的频率宽度
feat=numpy.dot(spec_power,fb.T) #对滤波器和能量谱进行点乘
feat=numpy.where(feat==0,numpy.finfo(float).eps,feat) #同样不能出现0
return feat,energy
def IDCTcoefficient(sig,
samplerate=16000,
win_length=0.025,
win_step=0.01,
pre_emphasis_coeff=0.97):
'''
计算初始IDCT系数
:param sig:
:param samplerate:
:param win_length:
:param win_step:
:param pre_emphasis_coeff:
:return:
'''
#预处理
signal = pre_emphasis(sig, pre_emphasis_coeff)
#分帧
frames = audio2frame(signal, win_length * samplerate,
win_step * samplerate) # 得到帧数组
#加窗
frames *= np.hamming(int(round(win_length * samplerate))) # 加窗
#DCT
dctfeat = dct(frames, type=2, axis=1, norm='ortho') # 进行离散余弦变换
#为0无法取对数
#dctfeat[dctfeat < 1e-30] = 1e-30
dctfeatnoz = np.where(dctfeat == 0,
np.finfo(float).eps,
dctfeat) #对为0的地方调整为eps，这样便于进行对数处理
dctfeatnoz = np.absolute(dctfeatnoz)
#取对数
logfeat = np.log(dctfeatnoz)
del dctfeatnoz
#取IDCT
idctfeat = idct(logfeat, type=2, axis=1, norm='ortho') # 进行反离散余弦变换
return idctfeat
