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)
def test_lmfcc(self):
lmfcc = lifter(self.example['mfcc'])
assert_allclose(lmfcc, self.example['lmfcc'])
# 4.5: Mel filterbank log spectrum
print("Applying Mel filterbank log spectrum...")
MSPEC = logMelSpectrum(FFT, 20000)
if compare(MSPEC, example['mspec']):
print("The result matches the example.")
else:
print("The result doesn't match the example.")
# 4.6: Cosine transform and liftering
print("Applying cosine transform...")
MFCC = cepstrum(MSPEC, 13)
if compare(MFCC, example['mfcc']): print("The result matches the example.")
else: print("The result doesn't match the example.")
print("Applying liftering...")
LMFCC = tools.lifter(MFCC)
if compare(LMFCC, example['lmfcc']):
print("The result matches the example.")
else:
print("The result doesn't match the example.")
# Apply to data
# First step
mfcc_features = mfcc(data[0]['samples']) # liftered
mspec_features = mspec(data[0]['samples'])
# Compute the mspec and mfcc for all utterances and stack them
for i in range(1, len(data)):
mfcc_features = np.vstack(
(mfcc_features, mfcc(data[i]['samples']))) # liftered
mspec_features = np.vstack((mspec_features, mspec(data[i]['samples'])))
pre_emph = preemp(window_frames)
if (np.array_equal(example['preemph'], pre_emph)):
customPlot(pre_emph, 'preemph: preemphasis', True)
windowed = windowing(pre_emph)
if (np.allclose(example['windowed'], windowed, atol=1e-08)):
customPlot(windowed, 'windowed: hamming window', True)
fft_ = powerSpectrum(windowed, 512)
if (np.allclose(example['spec'], fft_, atol=1e-08)):
customPlot(fft_, 'spec:abs(FFT)^2', True)
logMel = logMelSpectrum(fft_, sampling_rate, 512)
if (np.allclose(example['mspec'], logMel, atol=1e-08)):
customPlot(logMel, 'mspec:Mel Filterbank', True)
mfcc_ = cepstrum(logMel, 13)
if (np.allclose(example['mfcc'], mfcc_, atol=1e-08)):
customPlot(mfcc_, 'mfcc:MFCCs', True)
lmfcc_ = lifter(mfcc_)
if (np.allclose(example['lmfcc'], lmfcc_, atol=1e-08)):
customPlot(lmfcc_, 'lmfcc:Liftered MFCCs', True)
from lab1_proto import mfcc, mspec
data = np.load('lab1_data.npz', allow_pickle=True)['data']
for i in range(data.shape[0]):
samples = data[i]['samples']
s = mfcc(samples)
t = mspec(samples)
if (i == 0):
data_mfcc = s
data_mspec = t
else:
data_mfcc = np.append(data_mfcc, s, axis=0)
data_mspec = np.append(data_mspec, t, axis=0)
