def extractor(filename):
frameSize = 1024
hopSize = 512
fs = 44100
audio = ess.MonoLoader(filename=filename, sampleRate=fs)()
w = ess.Windowing(type='hamming', normalized=False)
# make sure these are same for MFCC and IDCT computation
NUM_BANDS = 26
DCT_TYPE = 2
LIFTERING = 0
NUM_MFCCs = 13
spectrum = ess.Spectrum()
mfcc = ess.MFCC(
numberBands=NUM_BANDS,
numberCoefficients=
NUM_MFCCs, # make sure you specify first N mfcc: the less, the more lossy (blurry) the smoothed mel spectrum will be
weighting=
'linear', # computation of filter weights done in Hz domain (optional)
normalize=
'unit_max', # htk filter normaliation to have constant height = 1 (optional)
dctType=DCT_TYPE,
logType='log',
liftering=LIFTERING) # corresponds to htk default CEPLIFTER = 22
idct = ess.IDCT(inputSize=NUM_MFCCs,
outputSize=NUM_BANDS,
dctType=DCT_TYPE,
liftering=LIFTERING)
all_melbands_smoothed = []
for frame in ess.FrameGenerator(audio,
frameSize=frameSize,
hopSize=hopSize):
spect = spectrum(w(frame))
melbands, mfcc_coeffs = mfcc(spect)
melbands_smoothed = np.exp(
idct(mfcc_coeffs)) # inverse the log taken in MFCC computation
all_melbands_smoothed.append(melbands_smoothed)
# transpose to have it in a better shape
# we need to convert the list to an essentia.array first (== numpy.array of floats)
# mfccs = essentia.array(pool['MFCC']).T
all_melbands_smoothed = essentia.array(all_melbands_smoothed).T
# and plot
plt.imshow(all_melbands_smoothed, aspect='auto',
interpolation='none') # ignore enery
# plt.imshow(mfccs, aspect = 'auto', interpolation='none')
plt.show() # unnecessary if you started "ipython --pylab"
def __init__(self, input_filename, fft_size, numMelBands):
fft_size_dummy = 1024
window_function_dummy = np.hanning
AudioProcessor.__init__(self, input_filename, fft_size_dummy,
window_function_dummy)
# self.inv_mfcc_transform = InvMFCC() # inverse mfcc transform
# self.inv_mfcc_transform.setup()
self.framesize = 2048 #
# self.framesize = 1102 # default frame size in htk, at rate of 44100
zeroPadding = fft_size - self.framesize
self.w = ess.Windowing(
type='hamming',
size=self.framesize,
zeroPadding=zeroPadding,
# normalized = False,
zeroPhase=False)
spectrumSize = fft_size // 2 + 1
self.spectrum = ess.Spectrum(size=fft_size)
self.mfcc = ess.MFCC(
inputSize=spectrumSize, # htk-like mfccs
type='magnitude',
warpingFormula='htkMel',
weighting='linear',
highFrequencyBound=8000,
lowFrequencyBound=0,
numberBands=numMelBands,
numberCoefficients=InvMFCCAudioProcessor.NUM_MFCC_COEFFS,
normalize='unit_max',
dctType=3,
logType='log',
liftering=22)
self.idct = ess.IDCT(inputSize=InvMFCCAudioProcessor.NUM_MFCC_COEFFS,
outputSize=numMelBands,
dctType=3,
liftering=22)