def mfcc(samples,
winlen=400,
winshift=200,
preempcoeff=0.97,
nfft=512,
nceps=13,
samplingrate=20000,
liftercoeff=22,
cepstrum_flag=True):
"""Computes Mel Frequency Cepstrum Coefficients.
Args:
samples: array of speech samples with shape (N,)
winlen: lenght of the analysis window
winshift: number of samples to shift the analysis window at every time step
preempcoeff: pre-emphasis coefficient
nfft: length of the Fast Fourier Transform (power of 2, >= winlen)
nceps: number of cepstrum coefficients to compute
samplingrate: sampling rate of the original signal
liftercoeff: liftering coefficient used to equalise scale of MFCCs
Returns:
N x nceps array with lifetered MFCC coefficients
"""
frames = enframe(samples, winlen, winshift)
preemph = preemp(frames, preempcoeff)
windowed = windowing(preemph)
spec = powerSpectrum(windowed, nfft)
mspec = logMelSpectrum(spec, samplingrate)
if cepstrum_flag is True:
ceps = cepstrum(mspec, nceps)
return tools.lifter(ceps, liftercoeff)
else:
return mspec
def mfcc(samples,
winlen=400,
winshift=200,
preempcoeff=0.97,
nfft=512,
nceps=13,
samplingrate=20000,
liftercoeff=22):
"""Computes Mel Frequency Cepstrum Coefficients.
Args:
samples: array of speech samples with shape (N,)
winlen: lenght of the analysis window
winshift: number of samples to shift the analysis window at every time step
preempcoeff: pre-emphasis coefficient
nfft: length of the Fast Fourier Transform (power of 2, >= winlen)
nceps: number of cepstrum coefficients to compute
samplingrate: sampling rate of the original signal
liftercoeff: liftering coefficient used to equalise scale of MFCCs
Returns:
N x nceps array with lifetered MFCC coefficients
"""
frames = enframe(samples, winlen, winshift)
#plt.rc('text',usetex=True)
#plt.pcolormesh(frames)
#plt.axis('off')
#plt.title('Enframed Samples',fontsize=20)
#plt.show()
preemph = preemp(frames, preempcoeff)
windowed = windowing(preemph)
spec = powerSpectrum(windowed, nfft)
mspec = logMelSpectrum(spec, samplingrate)
ceps = cepstrum(mspec, nceps)
#plt.pcolormesh(ceps)
#plt.title('Mfcc Coefficients',fontsize=20)
#plt.axis('off')
#plt.show()
lmfcc = to.lifter(ceps, liftercoeff)
#corr=np.corrcoef(lmfcc)
#plt.pcolormesh(corr)
#plt.show()
return lmfcc
def mfcc(samples,
winlen=400,
winshift=200,
preempcoeff=0.97,
nfft=512,
nceps=13,
samplingrate=20000,
liftercoeff=22):
"""Computes Mel Frequency Cepstrum Coefficients.
Args:
samples: array of speech samples with shape (N,)
winlen: lenght of the analysis window
winshift: number of samples to shift the analysis window at every time step
preempcoeff: pre-emphasis coefficient
nfft: length of the Fast Fourier Transform (power of 2, >= winlen)
nceps: number of cepstrum coefficients to compute
samplingrate: sampling rate of the original signal
liftercoeff: liftering coefficient used to equalise scale of MFCCs
Returns:
N x nceps array with lifetered MFCC coefficients
"""
mspec_ = mspec(samples, winlen, winshift, preempcoeff, nfft, samplingrate)
ceps = cepstrum(mspec_, nceps)
return lifter(ceps, liftercoeff)
# Compute 4.4 : Fast Fourier Transform plt.subplot(815) result4 = proto.powerSpectrum(result3, 512) plt.imshow(result4.transpose(), origin='lower', interpolation='nearest', aspect='auto') # Compute 4.5 : Mel filterbank log spectrum plt.subplot(816) result5 = proto.logMelSpectrum(result4, 20000) plt.imshow(result5.transpose(), origin='lower', interpolation='nearest', aspect='auto') # Compute 4.6 : Cosine Transform and Liftering plt.subplot(817) result6 = proto.cepstrum(result5, 13) plt.imshow(result6.transpose(), origin='lower', interpolation='nearest', aspect='auto') plt.subplot(818) result7 = tools.lifter(result6) plt.imshow(result7.transpose(), origin='lower', interpolation='nearest', aspect='auto') ## Compute 5 : MFCC for all tidigits and concanate them #tidiMfcc = tools.mfcc(tidigits[0]['samples']) #for i in range(1, len(tidigits)): # tidiMfcc = np.append(tidiMfcc, tools.mfcc(tidigits[i]['samples']), axis=0 ) # ## Correlation #corMfcc = np.corrcoef(tidiMfcc.transpose()) # #tidiMspec = tools.mspec(tidigits[0]['samples']) #for i in range(1, len(tidigits)): # tidiMspec = np.append(tidiMspec, tools.mspec(tidigits[i]['samples']), axis=0 ) # Correlation
import numpy as np
import matplotlib.pyplot as plt
import proto
from tools import lifter
reload(proto)
example = np.load('C:\Users\Linnea\Desktop\dt2118_lab1_2016-04-01\example.npz'
)['example'].item()
tidigits = np.load(
'C:\Users\Linnea\Desktop\dt2118_lab1_2016-04-01\etidigits.npz')['tidigits']
#plt.plot(example['samples'])
# SHOW IMAGE FROM EXAMPLE
#plt.imshow(example['frames'], origin = 'lower', interpolation = 'nearest', aspect = 'auto')
ef = proto.enframe(example['samples'], 400, 200)
preemphas = proto.preemp(example['frames'], 0.97)
fourierTrans = proto.powerSpectrum(example['windowed'], 512)
cos = proto.cepstrum(example['mspec'], 13)
cosLift = lifter(cos, 22)
#plt.imshow(ef, origin = 'lower', interpolation = 'nearest', aspect = 'auto')
#plt.show()
plt.subplot(816)
result5 = proto.logMelSpectrum(result4, 20000)
plt.imshow(result5.transpose(),
origin='lower',
interpolation='nearest',
aspect='auto')
# Compute 4.6 : Cosine Transform and Liftering
plt.subplot(817)
result6 = proto.cepstrum(result5, 13)
plt.imshow(result6.transpose(),
origin='lower',
interpolation='nearest',
aspect='auto')
plt.subplot(818)
result7 = tools.lifter(result6)
plt.imshow(result7.transpose(),
origin='lower',
interpolation='nearest',
aspect='auto')
## Compute 5 : MFCC for all tidigits and concanate them
#tidiMfcc = tools.mfcc(tidigits[0]['samples'])
#for i in range(1, len(tidigits)):
# tidiMfcc = np.append(tidiMfcc, tools.mfcc(tidigits[i]['samples']), axis=0 )
#
## Correlation
#corMfcc = np.corrcoef(tidiMfcc.transpose())
#
#tidiMspec = tools.mspec(tidigits[0]['samples'])
#for i in range(1, len(tidigits)):
def main():
#------------------------------------------------ Quest 4 ----------------------------->>>
# Step 1: Enframe
sample_rate = example['samplingrate']
winlen = int(.02 * sample_rate) # window length = 20ms
winshift = int(.01 * sample_rate) # shift length = 10ms
enf = enframe(samples, winlen, winshift)
# Step 2: Pre-emphasis
pre_emp = preemp(enf, p=0.97)
# Step 3: Hamming Window
ham = windowing(pre_emp)
# Step 4: Fast Fourier Transform
FFT = powerSpectrum(ham, nfft=512)
# Step 5: Mel filterbank log spectrum
logMel = logMelSpectrum(FFT, sample_rate)
# Step 6: Cosine Transofrm and Liftering
nceps = 13
cos_tra = cepstrum(logMel, nceps)
l_cos_tra = tools.lifter(cos_tra)
# Tidigits test for Quest 4
l_MFCC = []
for item in tidigits:
tid_samples = item['samples']
l_MFCC.append(mfcc(tid_samples))
#------------------------------------------------ Quest 5 ----------------------------->>>
l_MFCC_concat = np.zeros([1, nceps])
for i in range(len(tidigits)):
l_MFCC_concat = np.append(l_MFCC_concat, l_MFCC[i], axis=0)
MFCC_cor_coef = np.corrcoef(
l_MFCC_concat, rowvar=0
) # rowvar = non-zero (default): each row represents a variable, with observations in the columns.
# Otherwise, the relationship is transposed: each column represents a variable, while the rows contain observations.
Mspec_concat = mspec(tidigits[0]['samples'])
for i in range(1, len(tidigits)):
Mspec_concat = np.append(Mspec_concat,
mspec(tidigits[i]['samples']),
axis=0)
Mspec_cor_coef = np.corrcoef(Mspec_concat, rowvar=0)
#------------------------------------------------ Quest 6 ----------------------------->>>
# check local dist function for 2 utterances from Tidigit samples
# utter1 = mfcc(tidigits[0]['samples'])
# utter2 = mfcc(tidigits[1]['samples'])
# loc_dist = locdist(utter1, utter2)
# next step
# D = np.zeros([len(tidigits), len(tidigits) ])
# N = len(D) # full data: len(D)
# M = D.shape[1] # full data: D.shape[1]
# for i in range(N):
# for j in range(M):
# u1 = mfcc(tidigits[i]['samples'])
# u2 = mfcc(tidigits[j]['samples'])
# loc_dist = locdist(u1, u2)
# D[i, j] = dtw(loc_dist)
#------------------------------------------------ Quest 7 ----------------------------->>>
all_features = np.copy(l_MFCC_concat)
n_components = 32
gmm = GMM(n_components=n_components, covariance_type='full')
feature_fit = gmm.fit(all_features)
lift_mfcc_71 = mfcc(tidigits[16]['samples'])
prediction_71 = gmm.predict(lift_mfcc_71)
print("prediction 71 =", prediction_71, "\n")
lift_mfcc_72 = mfcc(tidigits[17]['samples'])
prediction_72 = gmm.predict(lift_mfcc_72)
print("prediction 72 =", prediction_72, "\n")
lift_mfcc_73 = mfcc(tidigits[38]['samples'])
prediction_73 = gmm.predict(lift_mfcc_73)
print("prediction 73 =", prediction_73, "\n")
lift_mfcc_74 = mfcc(tidigits[39]['samples'])
prediction_74 = gmm.predict(lift_mfcc_74)
print("prediction 74 =", prediction_74, "\n")
lift_mfcc_un = mfcc(tidigits[4]['samples'])
prediction_un = gmm.predict(lift_mfcc_un)
print("prediction un =", prediction_un)
fig0 = plt.figure()
fig0.canvas.set_window_title('num of components = ' + str(n_components))
lab = tools.tidigit2labels(tidigits)
plt.plot(prediction_71, 'r', label=lab[16])
plt.legend(loc='upper right')
plt.suptitle('num of components = ' + str(n_components))
#plt.savefig('fig13')
fig1 = plt.figure()
fig1.canvas.set_window_title('num of components =' + str(n_components))
plt.plot(prediction_72, 'b', label=lab[17])
plt.legend(loc='upper right')
plt.suptitle('num of components = ' + str(n_components))
#plt.savefig('fig14')
fig2 = plt.figure()
fig2.canvas.set_window_title('num of components =' + str(n_components))
plt.plot(prediction_73, 'g', label=lab[38])
plt.legend(loc='upper right')
plt.suptitle('num of components = ' + str(n_components))
#plt.savefig('fig15')
fig3 = plt.figure()
fig3.canvas.set_window_title('num of components =' + str(n_components))
plt.plot(prediction_74, 'y', label=lab[39])
plt.legend(loc='upper right')
plt.suptitle('num of components = ' + str(n_components))
#plt.savefig('fig16')
#@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ VISUALIZATION @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
# # My Visualization:
# -------- Quest 4
fig1 = plt.figure(figsize=(10, 10))
fig1.canvas.set_window_title('My results')
# Speech samples
plt.subplot(8, 1, 1)
plt.plot(samples)
plt.title('Speech samples')
# Step 1: Enframe
plt.subplot(8, 1, 2)
plt.pcolormesh(enf.T, cmap='jet')
plt.title('Enframe')
# Step 2: Pre-emphasis
plt.subplot(8, 1, 3)
plt.pcolormesh(pre_emp.T, cmap='jet')
plt.title('Pre-emphasis')
# Step 3: Hamming Window
plt.subplot(8, 1, 4)
plt.pcolormesh(ham.T, cmap='jet')
plt.title('Hamming Window')
# Step 4: Fast Fourier Transform
plt.subplot(8, 1, 5)
plt.pcolormesh(FFT.T, cmap='jet')
plt.title('Fast Fourier Transform')
# Step 5: Mel filterbank log spectrum
plt.subplot(8, 1, 6)
plt.pcolormesh(logMel.T, cmap='jet')
plt.title('Mel filterbank log spectrum')
# Step 6: Cosine Transofrm and Liftering
plt.subplot(8, 1, 7)
plt.pcolormesh(cos_tra.T, cmap='jet')
plt.title('MFCC')
plt.subplot(8, 1, 8)
plt.pcolormesh(l_cos_tra.T, cmap='jet')
plt.title('Lifted MFCC')
# filter banks plot
fig2 = plt.figure()
fig2.canvas.set_window_title('Filter Banks')
plt.plot(fbank.T)
# Quest 4.6 : Tidigit utterances
fig3 = plt.figure(figsize=(10, 6))
fig3.canvas.set_window_title('Tidigit 10 utterances: l_MFCC')
plt.subplot(4, 1, 1)
plt.title(lab[16])
plt.xticks([])
plt.yticks([])
plt.pcolormesh(l_MFCC[16].T, cmap='jet')
plt.subplot(4, 1, 2)
plt.title(lab[17])
plt.xticks([])
plt.yticks([])
plt.pcolormesh(l_MFCC[17].T, cmap='jet')
plt.subplot(4, 1, 3)
plt.title(lab[38])
plt.xticks([])
plt.yticks([])
plt.pcolormesh(l_MFCC[38].T, cmap='jet')
plt.subplot(4, 1, 4)
plt.title(lab[39])
plt.xticks([])
plt.yticks([])
plt.pcolormesh(l_MFCC[39].T, cmap='jet')
fig3.tight_layout(pad=2)
# -------- Quest 5
fig4 = plt.figure(figsize=(10, 6))
fig4.canvas.set_window_title('Feature Correlation')
plt.subplot(2, 1, 1)
plt.pcolormesh(MFCC_cor_coef, cmap='jet')
plt.subplot(2, 1, 2)
plt.pcolormesh(Mspec_cor_coef, cmap='jet')
# -------- Quest 6
# plt.pcolormesh(D, cmap = 'jet')
# plt.show()
# fig5 = plt.figure(figsize=(20,30))
# fig5.canvas.set_window_title('Comparing Utterances')
# ax = fig5.add_subplot(111)
# link = hac.linkage(D, method = 'complete')
# labels = tools.tidigit2labels(tidigits)
# dendro = hac.dendrogram(link, labels = labels)
# ax.set_xticklabels(labels)
# ax.legend(labels)
# plt.legend(labels, loc='center', fontsize='xx-large')
fig1.tight_layout(pad=2)
fig2.tight_layout(pad=2)
plt.show() # uncomment for visualization