def super_vector(test_file_name, ubm_file):
wav = mywave()
waveData = wav.WaveRead(test_file_name)
waveVadIdx = vad(waveData**2)
waveData = waveData[waveVadIdx]
MFCC_obj = MFCC(40, 12, 300, 3400, 0.97, 16000, 50, 0.0256, 256)
MFCC_coef = MFCC_obj.sig2s2mfc(waveData)
ubm = GMM(n_mix=128, n_dim=12)
ubm.read(ubm_file)
ubm.adapt(MFCC_coef)
return ubm.means
def super_vector(test_file_name, ubm_file):
wav = mywave()
waveData = wav.WaveRead(test_file_name)
waveVadIdx = vad(waveData ** 2)
waveData = waveData[waveVadIdx]
MFCC_obj = MFCC(40,12,300,3400,0.97,16000,50,0.0256,256)
MFCC_coef = MFCC_obj.sig2s2mfc(waveData)
ubm = GMM(n_mix = 128, n_dim = 12)
ubm.read(ubm_file)
ubm.adapt(MFCC_coef)
return ubm.means
ubm_data_dirs = os.listdir(ubm_dir)
dim = 12
sig = np.array([])
features_M = np.ndarray(shape = (0,dim), dtype = 'float64')
features_F = np.ndarray(shape = (0,dim), dtype = 'float64')
features = np.ndarray(shape = (0,dim), dtype = 'float64')
wav = mywave()
print 'hello'
for ubm_data_dir in ubm_data_dirs:
print 'hello'
print ubm_data_dir
if ubm_data_dir == '.DS_Store':
continue
sig = wav.WaveRead(ubm_dir+r'/'+ubm_data_dir)
MFCC_obj = MFCC(40,12,300,3400,0.97,16000,50,0.0256,256)
MFCC_coef = MFCC_obj.sig2s2mfc(sig)
#energy = np.ndarray(shape = (MFCC_coef.shape[0],1),dtype = 'float64')
#energy[:,0] = 10*numpy.log10((MFCC_coef**2).sum(axis=1))
#MFCC_coef = np.hstack((MFCC_coef,energy))
"""
dtm1 = np.ndarray(shape = MFCC_coef.shape,dtype = 'float64' )
#初始化dtm1
dtm1[0:2,:] = 0
dtm1[MFCC_coef.shape[0]-2:MFCC_coef.shape[0],:] = 0;
#计算dtm1
for loop2 in range(2,MFCC_coef.shape[0]-2):
dtm1[loop2,:] = -2*MFCC_coef[loop2-2,:]-MFCC_coef[loop2-1,:]+MFCC_coef[loop2+1,:]+2*MFCC_coef[loop2+2,:]
dtm1 = dtm1/3;
dtm2 = np.ndarray(shape = MFCC_coef.shape,dtype = 'float64' )
#初始化dtm2
dtm2[0:4,:] = 0
ubm_data_dirs = os.listdir(ubm_dir)
dim = 12
sig = np.array([])
features_M = np.ndarray(shape=(0, dim), dtype='float64')
features_F = np.ndarray(shape=(0, dim), dtype='float64')
features = np.ndarray(shape=(0, dim), dtype='float64')
wav = mywave()
print 'hello'
for ubm_data_dir in ubm_data_dirs:
print 'hello'
print ubm_data_dir
if ubm_data_dir == '.DS_Store':
continue
sig = wav.WaveRead(ubm_dir + r'/' + ubm_data_dir)
MFCC_obj = MFCC(40, 12, 300, 3400, 0.97, 16000, 50, 0.0256, 256)
MFCC_coef = MFCC_obj.sig2s2mfc(sig)
#energy = np.ndarray(shape = (MFCC_coef.shape[0],1),dtype = 'float64')
#energy[:,0] = 10*numpy.log10((MFCC_coef**2).sum(axis=1))
#MFCC_coef = np.hstack((MFCC_coef,energy))
"""
dtm1 = np.ndarray(shape = MFCC_coef.shape,dtype = 'float64' )
#初始化dtm1
dtm1[0:2,:] = 0
dtm1[MFCC_coef.shape[0]-2:MFCC_coef.shape[0],:] = 0;
#计算dtm1
for loop2 in range(2,MFCC_coef.shape[0]-2):
dtm1[loop2,:] = -2*MFCC_coef[loop2-2,:]-MFCC_coef[loop2-1,:]+MFCC_coef[loop2+1,:]+2*MFCC_coef[loop2+2,:]
dtm1 = dtm1/3;
dtm2 = np.ndarray(shape = MFCC_coef.shape,dtype = 'float64' )
#初始化dtm2
dtm2[0:4,:] = 0
import mywave
from gmm import *
from MFCC import *
if __name__ == '__main__':
print __doc__
ubms_dir = 'ubms'
speaker_model_dir = 'adaption'
if not os.path.exists(speaker_model_dir):
os.mkdir(speaker_model_dir)
train_data_dir = 'train_data'
train_data = os.listdir(train_data_dir)
wav = mywave.mywave()
for train_wav in train_data:
print train_wav
wave_data = wav.WaveRead(train_data_dir + r'/' + train_wav)
MFCC_obj = MFCC(40, 12, 300, 3400, 0.97, 16000, 50, 0.0256, 256)
MFCC_coef = MFCC_obj.sig2s2mfc(wave_data)
adapted_gmm = GMM()
if train_wav[-5] == 'M':
adapted_gmm.read(ubms_dir + r'/ubm_M')
elif train_wav[-5] == 'F':
adapted_gmm.read(ubms_dir + r'/ubm_F')
else:
print 'train_wav name unexpected'
adapted_gmm.adapt(MFCC_coef)
adapted_gmm.write(speaker_model_dir + r'/' + train_wav)
if __name__ == '__main__': print __doc__ ubms_dir = 'ubms' speaker_model_dir = 'adaption' if not os.path.exists(speaker_model_dir): os.mkdir(speaker_model_dir) train_data_dir = 'train_data' train_data = os.listdir(train_data_dir) wav = mywave.mywave() for train_wav in train_data: print train_wav wave_data = wav.WaveRead(train_data_dir+r'/'+train_wav) MFCC_obj = MFCC(40,12,300,3400,0.97,16000,50,0.0256,256) MFCC_coef = MFCC_obj.sig2s2mfc(wave_data) adapted_gmm = GMM() if train_wav[-5] == 'M': adapted_gmm.read(ubms_dir+r'/ubm_M') elif train_wav[-5] == 'F': adapted_gmm.read(ubms_dir+r'/ubm_F') else: print 'train_wav name unexpected' adapted_gmm.adapt(MFCC_coef) adapted_gmm.write(speaker_model_dir+r'/'+train_wav)
