def FeatureTimeRms(x, iBlockLength, iHopLength, f_s):
# create blocks
xBlocks = pyACA.ToolBlockAudio(x, iBlockLength, iHopLength)
# number of results
iNumOfBlocks = xBlocks.shape[0]
# compute time stamps
t = (np.arange(0, iNumOfBlocks) * iHopLength + (iBlockLength / 2)) / f_s
# allocate memory
vrms = np.zeros(iNumOfBlocks)
for n, block in enumerate(xBlocks):
# calculate the rms
vrms[n] = np.sqrt(np.dot(block, block) / block.size)
# convert to dB
epsilon = 1e-5 # -100dB
vrms[vrms < epsilon] = epsilon
vrms = 20 * np.log10(vrms)
return vrms, t
def FeatureTimeAcfCoeff(x, iBlockLength, iHopLength, f_s, eta=19):
# create blocks
xBlocks = pyACA.ToolBlockAudio(x, iBlockLength, iHopLength)
# number of results
iNumOfBlocks = xBlocks.shape[0]
if (np.isscalar(eta)):
iNumOfResultsPerBlock = 1
else:
iNumOfResultsPerBlock = eta.size
# compute time stamps
t = (np.arange(0, iNumOfBlocks) * iHopLength + (iBlockLength / 2)) / f_s
# allocate memory
vacf = np.zeros([iNumOfResultsPerBlock, iNumOfBlocks])
for n, block in enumerate(xBlocks):
# calculate the acf
if not block.sum():
vacf[np.arange(0, iNumOfResultsPerBlock),
n] = np.zeros(iNumOfResultsPerBlock)
continue
else:
afCorr = np.correlate(block, block, "full") / np.dot(block, block)
# find the coefficients specified in eta
vacf[np.arange(0, iNumOfResultsPerBlock),
n] = afCorr[iBlockLength + eta]
return vacf, t
def FeatureTimeStd(x, iBlockLength, iHopLength, f_s):
# create blocks
xBlocks = pyACA.ToolBlockAudio(x, iBlockLength, iHopLength)
# number of results
iNumOfBlocks = xBlocks.shape[0]
# compute time stamps
t = (np.arange(0, iNumOfBlocks) * iHopLength + (iBlockLength / 2)) / f_s
# allocate memory
vstd = np.zeros(iNumOfBlocks)
for n, block in enumerate(xBlocks):
# calculate the rms
vstd[n] = np.std(block)
return vstd, t
def FeatureTimeZeroCrossingRate(x, iBlockLength, iHopLength, f_s):
# create blocks
xBlocks = pyACA.ToolBlockAudio(x, iBlockLength, iHopLength)
# number of results
iNumOfBlocks = xBlocks.shape[0]
# compute time stamps
t = (np.arange(0, iNumOfBlocks) * iHopLength + (iBlockLength / 2)) / f_s
# allocate memory
vzc = np.zeros(iNumOfBlocks)
for n, block in enumerate(xBlocks):
# calculate the zero crossing rate
vzc[n] = 0.5 * np.mean(np.abs(np.diff(np.sign(block))))
return vzc, t
def FeatureTimePeakEnvelope(x, iBlockLength, iHopLength, f_s):
# create blocks
xBlocks = pyACA.ToolBlockAudio(x, iBlockLength, iHopLength)
# number of results
iNumOfBlocks = xBlocks.shape[0]
# compute time stamps
t = (np.arange(0, iNumOfBlocks) * iHopLength + (iBlockLength / 2)) / f_s
alpha = 1 - np.array(
[np.exp(-2.2 / (f_s * 0.01)),
np.exp(-2.2 / (f_s * 1.5))])
# allocate memory
vppm = np.zeros([2, iNumOfBlocks])
v_tmp = np.zeros(iBlockLength)
for n, block in enumerate(xBlocks):
x_block = np.abs(block)
# detect the maximum per block
vppm[0, n] = np.max(x_block)
# calculate the PPM value - take into account block overlaps
# and discard concerns wrt efficiency
v_tmp = ppm(x_block, v_tmp[iHopLength - 1], alpha)
vppm[1, n] = np.max(v_tmp)
# convert to dB
epsilon = 1e-5 # -100dB
vppm[vppm < epsilon] = epsilon
vppm = 20 * np.log10(vppm)
return vppm, t
def FeatureTimeMaxAcf(x, iBlockLength, iHopLength, f_s, f_max=2000, fMinThresh=0.35):
# create blocks
xBlocks = pyACA.ToolBlockAudio(x, iBlockLength, iHopLength)
# number of results
iNumOfBlocks = xBlocks.shape[0]
# compute time stamps
t = (np.arange(0, iNumOfBlocks) * iHopLength + (iBlockLength / 2)) / f_s
# allocate memory
vacf = np.zeros(iNumOfBlocks)
for n, block in enumerate(xBlocks):
eta_min = np.floor(f_s / f_max).astype(int)
# calculate the acf
if not block.sum():
continue
else:
afCorr = np.correlate(block, block, "full") / np.dot(block, block)
afCorr = afCorr[np.arange(iBlockLength, afCorr.size)]
# update eta_min to avoid main lobe
eta_tmp = np.argmax(afCorr < fMinThresh)
eta_min = np.max([eta_min, eta_tmp])
afDeltaCorr = np.diff(afCorr)
eta_tmp = np.argmax(afDeltaCorr > 0)
eta_min = np.max([eta_min, eta_tmp])
# find the coefficients specified in eta
vacf[n] = np.max(afCorr[np.arange(eta_min + 1, afCorr.size)])
return vacf, t