def MRCG(x, fs=44100, framesize1=0.02, framesize2=0.2, hopsize=0.01):
hopsize = int(hopsize * fs)
# spectrogram init
winAnalysis = 'hann'
####---- cochleagram 1
framesize = int(framesize1 * fs)
N = 2 * framesize # padding 1 time framesize
SPECTRUM = ess.Spectrum(size=N)
WINDOW = ess.Windowing(type=winAnalysis, zeroPadding=N - framesize)
highFrequencyBound = fs / 2 if fs / 2 < 11000 else 11000
ERBBANDS = ess.ERBBands(sampleRate=fs,
highFrequencyBound=highFrequencyBound,
inputSize=framesize + 1)
cochlea1 = []
for frame in ess.FrameGenerator(x, frameSize=framesize, hopSize=hopsize):
frame = WINDOW(frame)
mXFrame = SPECTRUM(frame)
erbFrame = np.log10(ERBBANDS(mXFrame) + np.finfo(np.float).eps)
cochlea1.append(erbFrame)
cochlea1 = np.array(cochlea1)
####---- cochleagram 2
framesize = int(framesize2 * fs)
N = 2 * framesize # padding 1 time framesize
SPECTRUM = ess.Spectrum(size=N)
WINDOW = ess.Windowing(type=winAnalysis, zeroPadding=N - framesize)
highFrequencyBound = fs / 2 if fs / 2 < 11000 else 11000
ERBBANDS = ess.ERBBands(sampleRate=fs,
highFrequencyBound=highFrequencyBound,
inputSize=framesize + 1)
cochlea2 = []
for frame in ess.FrameGenerator(x, frameSize=framesize, hopSize=hopsize):
frame = WINDOW(frame)
mXFrame = SPECTRUM(frame)
erbFrame = np.log10(ERBBANDS(mXFrame) + np.finfo(np.float).eps)
cochlea2.append(erbFrame)
cochlea2 = np.array(cochlea2)
####---- smoothed version
cochlea3 = get_avg(cochlea1, 5, 5)
cochlea4 = get_avg(cochlea1, 11, 11)
all_cochleas = np.hstack((cochlea1, cochlea2, cochlea3, cochlea4))
####---- delta
d_all_cochleas = Fdeltas(all_cochleas.T)
dd_all_cochleas = Fdeltas(Fdeltas(all_cochleas.T, 5), 5)
d_all_cochleas = d_all_cochleas.T
dd_all_cochleas = dd_all_cochleas.T
return all_cochleas, d_all_cochleas, dd_all_cochleas
def _key_fnc(
sample: NDArray[Float32],
frequency_rate: int,
windowfnc: Window,
key_type: KeyFunction,
):
"""
This function computes the key function,
which in return calculates the keys for the [this.samples] map.
To calculate the spectral centroid,
the frequency_rate should be equal to the half of the samplerate.
"""
if key_type == KeyFunction.CENTROID:
return _get_centroid(
sample,
estd.Centroid(range=frequency_rate),
estd.Spectrum(),
estd.Windowing(type=windowfnc.value),
)
if key_type == KeyFunction.MAX:
return _get_max(
sample,
estd.Spectrum(),
estd.Windowing(type=windowfnc.value),
)
if key_type == KeyFunction.MFCC:
return _get_mfcc(
sample,
estd.MFCC(),
estd.Spectrum(),
estd.Windowing(type=windowfnc.value),
)
if key_type == KeyFunction.MELBANDS:
return _get_melbands(
sample,
estd.MFCC(),
estd.Spectrum(),
estd.Windowing(type=windowfnc.value),
)
if key_type == KeyFunction.MELBANDS_LOG:
return estd.UnaryOperator(type="log")(_get_melbands(
sample,
estd.MFCC(),
estd.Spectrum(),
estd.Windowing(type=windowfnc.value),
))
raise ValueError("Keyfunction is not defined!")
def extract(fname, outpath, fs=22050, fsize=1024, hsize=512):
"""
extract(fname, outpath, fs, fsize, hsize) will compute the mfcc of Audio file fname.
Inputs:
fname -- is the name of audio file.
outpath -- is the output path of processed files.
fs -- is the sampling frequency (Hz).
fsize -- is the size of each frame.
hsize -- is the hop size betwean frames.
Outputs:
the file contains the mfcc coefficents of audio file.
in what format???
"""
# gate(fname)
loader = es.MonoLoader(filename=fname, sampleRate=fs)
# length = len(loader)
# maxim = max(loader)
# for sample in loader:
# if abs(sample) < maxim/20:
# sample = 0 ;
w = es.Windowing(type='hann')
spectrum = es.Spectrum()
mfcc = es.MFCC(inputSize=513, numberCoefficients=20)
mfccs = []
audio = loader()
for frame in es.FrameGenerator(audio, frameSize=1024, hopSize=512):
mfcc_bands, mfcc_coeffs = mfcc(spectrum(w(frame)))
mfccs.append(mfcc_coeffs)
mfccs = np.array(mfccs)
return mfcc
def vibFreq(pitchtrack, sp, hopsize):
'''
:param pitchtrack:
:param sp: samplerate of wave audio
:param hopsize:
:return: 3 frequencies of potential vibrato
'''
if pitchtrack.dtype != np.float32:
pitchtrack = pitchtrack.astype(np.float32)
pitchtrackPad = pitchtrack[:]
sampleRate = sp / hopsize
ptlen = len(pitchtrack)
fftSize = int(pow(2, ceil(log(ptlen) /
log(2)))) # next pow of pitchtrack length
if ptlen < fftSize:
pitchtrackPad = np.append(pitchtrack,
np.zeros(fftSize - ptlen, dtype=np.float32))
S = ess.Spectrum(size=fftSize)(pitchtrackPad)
locs, amps = ess.PeakDetection(maxPeaks=3, orderBy='amplitude')(S)
freqs = locs * (fftSize / 2 + 1) * sampleRate / fftSize
return freqs[0]
def segment(audio, hopSize, frameSize, rms_onset_threshold,
mel_onset_threshold, flux_onset_threshold, onset_threshold):
# init algorithms
o_mel = estd.OnsetDetection(method='melflux')
o_rms = estd.OnsetDetection(method='rms')
o_hfc = estd.OnsetDetection(method='hfc')
o_flux = estd.OnsetDetection(method='flux')
fft = estd.FFT()
c2p = estd.CartesianToPolar()
pool = essentia.Pool()
frame_generator = estd.FrameGenerator(audio,
frameSize=frameSize,
hopSize=hopSize)
w = estd.Windowing(type='hann')
yin = estd.PitchYinFFT(frameSize=frameSize,
minFrequency=40,
maxFrequency=2500,
interpolate=True)
spectrum = estd.Spectrum()
loudness = estd.Loudness()
# control parameters
attack = False
detection = True
mel_onset_value = 0
rms_onset_value = 0
# output variables
onset = None
sustain = None
for index, frame in enumerate(frame_generator):
mag, phase = c2p(fft(w(frame)))
_, conf = yin(spectrum(w(frame)))
loud = loudness(frame)
mel_onset = o_mel(mag, phase)
rms_onset = o_rms(mag, phase)
hfc_onset = o_hfc(mag, phase)
flux_onset = o_flux(mag, phase)
pool.add('onsets_mel', mel_onset)
pool.add('onsets_rms', rms_onset)
pool.add('onsets_hfc', hfc_onset)
pool.add('onsets_flux', flux_onset)
pool.add('conf', conf)
pool.add('loudness', loud)
# condition for onset
if detection and (flux_onset > flux_onset_threshold or mel_onset > mel_onset_threshold) \
and rms_onset > rms_onset_threshold and loud > onset_threshold:
onset = index
attack = True
detection = False
mel_onset_value = mel_onset
rms_onset_value = rms_onset
# condition for beginning of sustain
if attack and conf > 0.5 and rms_onset < rms_onset_value * .05 and mel_onset < mel_onset_value * .3:
attack = False
sustain = index
return onset, sustain
def compute_description(x,
M=WINDOW_SIZE,
N=FFT_SIZE,
H=HOP_SIZE,
fs=SR,
window_type=WINDOW_TYPE):
'''
-extract features from audio file
-Features:
HFC
SPECTRAL CENTROID
SPECTRAL ENERGY
F0
loud_factor = energy * (spectral_centroid - F0) #how many harmonics = how much speaker is yelling
PITCH CONFIDENCE
'''
#audioLoader = ess.EasyLoader(filename=file_name, sampleRate=fs)
#create essentia instances
x = essentia.array(x)
spectrum = ess.Spectrum(size=N)
window = ess.Windowing(size=M, type=window_type)
hfc = ess.HFC(sampleRate=fs)
spectralCentroid = ess.SpectralCentroidTime(sampleRate=fs)
energy = ess.Energy()
pitch_extractor = ess.PredominantPitchMelodia(frameSize=M,
hopSize=H,
maxFrequency=1200)
#init vectors
CONTRAST = []
HFC = []
CENTROID = []
ENERGY = []
#compute features for every stft frame
for frame in ess.FrameGenerator(x,
frameSize=M,
hopSize=H,
startFromZero=True): #generate frames
wX = window(frame) #window frame
mX = spectrum(wX) #compute fft
frame_hfc = hfc(mX)
HFC.append(frame_hfc)
frame_centroid = spectralCentroid(
wX) #compute spectral centroid in time domain
CENTROID.append(frame_centroid)
frame_energy = energy(mX) #compute spectral energy in time domain
ENERGY.append(frame_energy)
F0, SALIENCE = pitch_extractor(x) #estimate pitch in time domain
#convert into numpy matrices
HFC = essentia.array(HFC)
CENTROID = essentia.array(CENTROID)
ENERGY = essentia.array(ENERGY)
F0 = essentia.array(F0)
SALIENCE = essentia.array(SALIENCE)
F0 = F0[:len(CENTROID)]
SALIENCE = SALIENCE[:len(CENTROID)]
return HFC, CENTROID, ENERGY, F0, SALIENCE
def extract_features(x,
M=WINDOW_SIZE,
N=FFT_SIZE,
H=HOP_SIZE,
fs=SR,
window_type=WINDOW_TYPE):
'''
extract magnitudes spectra from input vector and apply power-law compression
'''
#init functions and vectors
x = essentia.array(x)
spectrum = ess.Spectrum(size=N)
window = ess.Windowing(size=M, type=window_type)
SP = []
#compute STFT
for frame in ess.FrameGenerator(x,
frameSize=M,
hopSize=H,
startFromZero=True): #generate frames
wX = window(frame) #window frame
mX = spectrum(wX) #compute fft
###############################OPTIMIZATION[[[[[[[[[[[[[[]]]]]]]]]]]]]]
#DEPRECATED
#################################################
SP.append(mX)
SP = essentia.array(SP)
SP = np.power(SP, 2. / 3.) #power law compression
return SP
def get_f0(audio, minf0=20, maxf0=22050, cf=0.9, ws=2048, hs=256):
'''
Args:
audio (array): audio signal (output from MonoLoader)
minf0 (int): minimum allowed frequency
maxf0 (int): maximun allowed frequency
cf (float): confidence threshold (0 - 1)
ws (int): window size
hp (int): hop size
Returns:
f0 (array):
'''
# instantiate Essentia functions
w = es.Windowing(type='hann', zeroPadding=ws)
spec = es.Spectrum()
yin = es.PitchYinFFT(minFrequency=minf0, maxFrequency=maxf0, frameSize=ws)
# empty lists for f0 and confidence
f0 = []
conf = []
# iterate over frames
for frame in es.FrameGenerator(audio, frameSize=ws, hopSize=hs):
p, pc = yin(spec(w(frame)))
f0.append(p)
conf.append(pc)
# convert lists to np.arrays
f0 = np.array(f0)
conf = np.array(conf)
# return f0 over given confidence
f0[conf < cf] = 0
return f0
def getHPCPEssentia(XAudio, Fs, winSize, hopSize, squareRoot=False, NChromaBins=36, NHarmonics = 0):
"""
Wrap around the essentia library to compute HPCP features
:param XAudio: A flat array of raw audio samples
:param Fs: Sample rate
:param winSize: Window size of each STFT window
:param hopSize: Hop size between STFT windows
:param squareRoot: Do square root compression?
:param NChromaBins: How many chroma bins (default 36)
:returns H: An (NChromaBins x NWindows) matrix of all \
chroma windows
"""
import essentia
from essentia import Pool, array
import essentia.standard as ess
spectrum = ess.Spectrum()
window = ess.Windowing(size=winSize, type='hann')
spectralPeaks = ess.SpectralPeaks()
hpcp = ess.HPCP(size=NChromaBins, harmonics=NHarmonics)
H = []
for frame in ess.FrameGenerator(array(XAudio), frameSize=winSize, hopSize=hopSize, startFromZero=True):
S = spectrum(window(frame))
freqs, mags = spectralPeaks(S)
H.append(hpcp(freqs, mags))
H = np.array(H)
H = H.T
if squareRoot:
H = sqrtCompress(H)
return H
def calc_chromagram(self):
# save the results in the stft_pool
self.chromagram = []
hpcp = es.HPCP(
size=12, # we will need higher resolution for Key estimation
referenceFrequency=440, # assume tuning frequency is 44100.
bandPreset=False,
weightType='cosine',
nonLinear=False,
windowSize=1.,
sampleRate=self.sample_rate)
spectrum = es.Spectrum(size=self.fft_size)
spectral_peaks = es.SpectralPeaks(sampleRate=self.sample_rate)
for frame in es.FrameGenerator(self.audio,
frameSize=self.frame_size,
hopSize=self.hop_size,
startFromZero=True):
frame = array(frame * self.window)
freqs, mags = spectral_peaks(spectrum(frame))
chroma = hpcp(freqs, mags)
self.chromagram.append(chroma)
self.chromagram = array(self.chromagram)
self.timeAxSec = np.arange(len(
self.chromagram)) * self.hop_size / float(self.sample_rate)
def extract_features(x,
M=Config.WINDOW_SIZE,
N=Config.FFT_SIZE,
H=Config.HOP_SIZE,
fs=Config.FS,
window_type=Config.WINDOW_TYPE):
'''
Function that extracts spectrogram from an audio signal
-----------------------
Input: Samples, window size (int), FFT size (int), Hop size (int),
Sampling rate, Window type (e.g. Hanning)
Output: Spectrogram
-----------------------
'''
# init functions and vectors
x = essentia.array(x)
spectrum = ess.Spectrum(size=N)
window = ess.Windowing(size=M, type=window_type)
SP = []
# compute STFT
for frame in ess.FrameGenerator(x,
frameSize=M,
hopSize=H,
startFromZero=True): # generate frames
wX = window(frame) # window frame
mX = spectrum(wX) # compute fft
SP.append(mX)
SP = essentia.array(SP)
SP = np.power(SP, 2. / 3.) # power law compression
SP = SP[:, :int(Config.FFT_SIZE / 4 + 1)]
return SP
def get_beat_chunks(filename, bpm_restrict=None):
audio = std.MonoLoader(filename=filename)()
hpcp = std.HPCP()
spectrum = std.Spectrum()
speaks = std.SpectralPeaks()
large_speaks = std.SpectralPeaks(maxPeaks=2000)
tivs = []
sr = 44100
bpm = get_tempo(filename)
tivs_framewise = []
if bpm_restrict != None and bpm_restrict != bpm:
raise ValueError
sec_beat = (60 / bpm)
beats = np.arange(0, len(audio) / sr, sec_beat)
beats = np.append(beats, len(audio) / sr)
for i in range(1, len(beats)):
segmented_audio = audio[int(beats[i - 1] * sr):int(beats[i] * sr)]
cutter = std.FrameGenerator(segmented_audio)
for sec in cutter:
spec = spectrum(sec)
freq, mag = speaks(spec)
chroma = hpcp(freq, mag)
tivs_framewise.append(chroma)
np2_seg_audio = zeropad_next_power_2(segmented_audio)
spec = spectrum(np2_seg_audio)
freq, mag = speaks(spec)
chroma = hpcp(freq, mag)
tivs.append(chroma)
# Calculate the whole TIV
np2_whole = zeropad_next_power_2(audio)
spec = spectrum(np2_whole)
freq, mag = large_speaks(spec)
chroma_whole = hpcp(freq, mag)
return mt.TIVCollection.from_pcp(np.array(tivs).T), mt.TIV.from_pcp(chroma_whole), mt.TIVCollection.from_pcp(np.array(tivs_framewise).T)
def getMBE(audio):
'''
mel band energy feature
:param audio:
:return:
'''
winAnalysis = 'hann'
# this MFCC is for pattern classification, which numberBands always be by default
MFCC40 = ess.MFCC(sampleRate=fs,
highFrequencyBound=highFrequencyBound,
inputSize=framesize + 1)
N = 2 * framesize # padding 1 time framesize
SPECTRUM = ess.Spectrum(size=N)
WINDOW = ess.Windowing(type=winAnalysis, zeroPadding=N - framesize)
mfccBands = []
for frame in ess.FrameGenerator(audio,
frameSize=framesize,
hopSize=hopsize):
frame = WINDOW(frame)
mXFrame = SPECTRUM(frame)
bands, mfccFrame = MFCC40(mXFrame)
mfccBands.append(bands)
feature = np.array(mfccBands)
return feature
def get_constantq(frames, sample_rate=16000, num_bands=64):
max_freq = 8000
min_freq = 125
num_octaves = np.log2(max_freq / min_freq)
bins_per_octave = int(np.ceil(num_bands / num_octaves))
frame_size = len(frames[0])
const_q_spectra = []
spectrum_estimator = es.Spectrum(size=frame_size)
if num_bands == 16:
padding_size = max([0, 512 - frame_size])
elif num_bands == 32:
padding_size = max([0, 2048 - frame_size])
else:
padding_size = max([0, 1024 - frame_size])
windowing = es.Windowing(type='hann',
size=frame_size,
zeroPadding=padding_size)
constantq_estimator = es.ConstantQ(binsPerOctave=bins_per_octave,
minFrequency=min_freq,
numberBins=num_bands,
sampleRate=sample_rate)
for frame in frames:
const_q_spectrum = constantq_estimator(windowing(frame))
const_q_spectra.append(np.abs(const_q_spectrum))
return np.array(const_q_spectra).T
def extract_features(x,
M=WINDOW_SIZE,
N=FFT_SIZE,
H=HOP_SIZE,
fs=SR,
window_type=WINDOW_TYPE):
'''
extract magnitudes spectra from input vector
apply power-law compression
cutt the upper spectrum
'''
#init functions and vectors
x = essentia.array(x)
spectrum = ess.Spectrum(size=N)
window = ess.Windowing(size=M, type=WINDOW_TYPE)
SP = []
#compute STFT
for frame in ess.FrameGenerator(x,
frameSize=M,
hopSize=H,
startFromZero=True): #generate frames
wX = window(frame) #window frame
mX = spectrum(wX) #compute fft
SP.append(mX)
SP = essentia.array(SP)
SP = np.power(SP, 2. / 3.) #power law compression
#SP = SP[:,:int(FFT_SIZE/2+1)] #cut upper spectrum (above 4 khz)
return SP
def file_to_hpcp(loop):
loop = e.array(loop)
windowing = es.Windowing(type='blackmanharris62')
spectrum = es.Spectrum()
spectral_peaks = es.SpectralPeaks(orderBy='magnitude',
magnitudeThreshold=0.001,
maxPeaks=20,
minFrequency=20,
maxFrequency=8000)
hpcp = es.HPCP(maxFrequency=8000)
spec_group = []
hpcp_group = []
for frame in es.FrameGenerator(loop, frameSize=1024, hopSize=512):
windowed = windowing(frame)
fft = spectrum(windowed)
frequencies, magnitudes = spectral_peaks(fft)
final_hpcp = hpcp(frequencies, magnitudes)
spec_group.append(fft)
hpcp_group.append(final_hpcp)
mean_hpcp = np.mean(np.array(hpcp_group).T, axis=1)
#normalize to 1
mean_hpcp = mean_hpcp / mean_hpcp.max()
return mean_hpcp
def _extract_pitch_contours(self, audio):
# Hann window with x4 zero padding
run_windowing = estd.Windowing( # pylint: disable-msg=E1101
zeroPadding=3 * self.frame_size)
run_spectrum = estd.Spectrum( # pylint: disable-msg=E1101
size=self.frame_size * 4)
run_spectral_peaks = estd.SpectralPeaks( # pylint: disable-msg=E1101
minFrequency=self.min_frequency,
maxFrequency=self.max_frequency,
magnitudeThreshold=self.magnitude_threshold,
sampleRate=self.sample_rate,
orderBy='magnitude')
# convert unit to cents, PitchSalienceFunction takes 55 Hz as the
# default reference
run_pitch_salience_function = \
estd.PitchSalienceFunction( # pylint: disable-msg=E1101
binResolution=self.bin_resolution)
run_pitch_salience_function_peaks = \
estd.PitchSalienceFunctionPeaks( # pylint: disable-msg=E1101
binResolution=self.bin_resolution,
minFrequency=self.min_frequency,
maxFrequency=self.max_frequency)
run_pitch_contours = estd.PitchContours( # pylint: disable-msg=E1101
hopSize=self.hop_size,
binResolution=self.bin_resolution,
peakDistributionThreshold=self.peak_distribution_threshold)
# compute frame by frame
pool = Pool()
for frame in estd.FrameGenerator(
audio, # pylint: disable-msg=E1101
frameSize=self.frame_size,
hopSize=self.hop_size):
frame = run_windowing(frame)
spectrum = run_spectrum(frame)
peak_frequencies, peak_magnitudes = run_spectral_peaks(spectrum)
salience = run_pitch_salience_function(peak_frequencies,
peak_magnitudes)
salience_peaks_bins, salience_peaks_contour_saliences = \
run_pitch_salience_function_peaks(salience)
if not np.size(salience_peaks_bins):
salience_peaks_bins = np.array([0])
if not np.size(salience_peaks_contour_saliences):
salience_peaks_contour_saliences = np.array([0])
pool.add('allframes_salience_peaks_bins', salience_peaks_bins)
pool.add('allframes_salience_peaks_contourSaliences',
salience_peaks_contour_saliences)
# post-processing: contour tracking
contours_bins, contour_saliences, contours_start_times, duration = \
run_pitch_contours(
[f.tolist()
for f in pool['allframes_salience_peaks_bins']],
[f.tolist()
for f in pool['allframes_salience_peaks_contourSaliences']])
return contours_bins, contours_start_times, contour_saliences, duration
def extractor(filename):
fs = 44100
audio = ess.MonoLoader(filename=filename, sampleRate=fs)()
# dynamic range expansion as done in HTK implementation
audio = audio * 2**15
frameSize = 1102 # corresponds to htk default WINDOWSIZE = 250000.0
hopSize = 441 # corresponds to htk default TARGETRATE = 100000.0
fftSize = 2048
spectrumSize = fftSize // 2 + 1
zeroPadding = fftSize - frameSize
w = ess.Windowing(
type='hamming', # corresponds to htk default USEHAMMING = T
size=frameSize,
zeroPadding=zeroPadding,
normalized=False,
zeroPhase=False)
spectrum = ess.Spectrum(size=fftSize)
mfcc_htk = ess.MFCC(
inputSize=spectrumSize,
type='magnitude', # htk uses mel filterbank magniude
warpingFormula='htkMel', # htk's mel warping formula
weighting='linear', # computation of filter weights done in Hz domain
highFrequencyBound=8000, # corresponds to htk default
lowFrequencyBound=0, # corresponds to htk default
numberBands=26, # corresponds to htk default NUMCHANS = 26
numberCoefficients=13,
normalize=
'unit_max', # htk filter normaliation to have constant height = 1
dctType=3, # htk uses DCT type III
logType='log',
liftering=22) # corresponds to htk default CEPLIFTER = 22
mfccs = []
# startFromZero = True, validFrameThresholdRatio = 1 : the way htk computes windows
for frame in ess.FrameGenerator(audio,
frameSize=frameSize,
hopSize=hopSize,
startFromZero=True,
validFrameThresholdRatio=1):
spect = spectrum(w(frame))
mel_bands, mfcc_coeffs = mfcc_htk(spect)
mfccs.append(mfcc_coeffs)
# 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
mfccs = essentia.array(mfccs).T
# and plot
plt.imshow(mfccs[1:, :], aspect='auto',
interpolation='none') # ignore enery
# plt.imshow(mfccs, aspect = 'auto', interpolation='none')
plt.show() # unnecessary if you started "ipython --pylab"
def essentiaObjectInit(self):
winAnalysis = 'hann'
self.MFCC80 = ess.MFCC(sampleRate=self.fs,
highFrequencyBound=self.highFrequencyBound,
inputSize=self.frameSize + 1,
numberBands=self.numberBands)
N = 2 * self.frameSize # padding 1 time framesize
self.SPECTRUM = ess.Spectrum(size=N)
self.WINDOW = ess.Windowing(type=winAnalysis, zeroPadding=N - self.frameSize)
def __onset_candidate_detection__(self):
spectrum = e.Spectrum()
e_onsetdetection = e.OnsetDetection(method="flux")
onsetspecs = []
for frame in e.FrameGenerator(self.signal, 1024, 512):
self.frames.append(frame)
onsetspecs.append(spectrum(frame))
self.onset_candidates.append(e_onsetdetection(onsetspecs[-1], [0]*len(onsetspecs[-1])))
self.frame_count = len(self.frames)
def __init__(self, frame_size, hop_size, window_type, feature,
beats, sample_rate):
"""STFTFeature constructor."""
self.frame_size = frame_size
self.hop_size = hop_size
self.window_type = window_type
self.w = ES.Windowing(type=window_type)
self.spectrum = ES.Spectrum()
self.feature = feature # Essentia feature object
self.beats = beats
self.sample_rate = sample_rate
def exp_env(audio, step):
def func(x, a, c):
return a * np.exp(-c * x)
max_pos = np.argmax(audio)
audio1 = audio
audio = audio[np.argmax(audio):]
step = int(step)
audio = np.abs(audio)
envelope = []
env_x = []
for i in range(0, len(audio), step):
env_x += [i + np.argmax(audio[i:i + step])]
envelope += [np.max(audio[i:i + step])]
env_x = np.array(env_x)
envelope = np.array(envelope)
try:
popt, pcov = curve_fit(func, env_x, envelope, p0=(1, 1e-3))
except RuntimeError:
popt = [envelope[0], 0]
pcov = []
xx = np.arange(0, len(audio), 1)
yy = func(xx, *popt)
xx = xx + max_pos
xx = np.append(np.arange(0, max_pos), xx)
yy = np.append(np.zeros(max_pos), yy)
plt.plot(xx, yy)
plt.plot(xx, audio1, color='green')
start = env_x[np.where(envelope == envelope.max())[0]]
nf1 = envelope[0:5].mean()
locs = np.where(envelope < 0.1 * envelope.max())[0]
if len(locs) < 1:
stop1 = env_x[-1]
else:
stop1 = env_x[locs[np.where(
locs > np.where(envelope == envelope.max())[0])][0]]
locs = np.where(envelope < 0.01 * envelope.max())[0]
if len(locs) < 1:
stop2 = env_x[-1]
else:
stop2 = env_x[locs[np.where(
locs > np.where(envelope == envelope.max())[0])][0]]
plt.xlabel('Samples')
plt.ylabel('Absolute Amplitude')
plt.axis([0, 140000, 0, 0.20])
plt.figure()
en_mod = np.array(audio1 - yy, dtype='float32')
if len(en_mod) % 2 > 0:
en_mod = en_mod[:-1]
spectrum = estd.Spectrum()(en_mod)
plt.plot(spectrum)
plt.show()
return stop1 - start
def mfcc_htk(self, window_length=22050, nmfcc=13, n_mels=26, fmax=8000, lifterexp=22):
"""
Get MFCCs 'the HTK way' with the help of Essentia
https://github.com/MTG/essentia/blob/master/src/examples/tutorial/example_mfcc_the_htk_way.py
Using all of the default parameters from there except the hop length (which shouldn't matter), and a much longer window length (which has been found to work better for covers)
Parameters
----------
window_length: int
Length of the window to use for the STFT
nmfcc: int
Number of MFCC coefficients to compute
n_mels: int
Number of frequency bands to use
fmax: int
Maximum frequency
Returns
-------
ndarray(nmfcc, nframes)
An array of all of the MFCC frames
"""
fftlen = int(2**(np.ceil(np.log(window_length)/np.log(2))))
spectrumSize= fftlen//2+1
zeroPadding = fftlen - window_length
w = estd.Windowing(type = 'hamming', # corresponds to htk default USEHAMMING = T
size = window_length,
zeroPadding = zeroPadding,
normalized = False,
zeroPhase = False)
spectrum = estd.Spectrum(size=fftlen)
mfcc_htk = estd.MFCC(inputSize = spectrumSize,
type = 'magnitude', # htk uses mel filterbank magniude
warpingFormula = 'htkMel', # htk's mel warping formula
weighting = 'linear', # computation of filter weights done in Hz domain
highFrequencyBound = fmax, # 8000 is htk default
lowFrequencyBound = 0, # corresponds to htk default
numberBands = n_mels, # corresponds to htk default NUMCHANS = 26
numberCoefficients = nmfcc,
normalize = 'unit_max', # htk filter normaliation to have constant height = 1
dctType = 3, # htk uses DCT type III
logType = 'log',
liftering = lifterexp) # corresponds to htk default CEPLIFTER = 22
mfccs = []
# startFromZero = True, validFrameThresholdRatio = 1 : the way htk computes windows
for frame in estd.FrameGenerator(self.audio_vector, frameSize = window_length, hopSize = self.hop_length , startFromZero = True, validFrameThresholdRatio = 1):
spect = spectrum(w(frame))
mel_bands, mfcc_coeffs = mfcc_htk(spect)
mfccs.append(mfcc_coeffs)
return np.array(mfccs, dtype=np.float32).T
def audio_features(audio_win):
"""
returns audio features for a win
"""
if audio_win.shape[0] % 2 == 1:
audio_win = audio_win[:-1]
spectrum = esst.Spectrum(size=audio_win.shape[0])(audio_win)
_bands, mfcc = esst.MFCC(inputSize=spectrum.shape[0],
sampleRate=SR)(spectrum)
rhythm = esst.RhythmDescriptors()(audio_win)
return mfcc.tolist() + [rhythm[2]] + list(rhythm[5:11])
def melspectrogram(audio,
sampleRate=44100,
frameSize=2048,
hopSize=1024,
window='blackmanharris62',
zeroPadding=0,
center=True,
numberBands=[128, 96, 48, 32, 24, 16, 8],
lowFrequencyBound=0,
highFrequencyBound=None,
weighting='linear',
warpingFormula='slaneyMel',
normalize='unit_tri'):
if highFrequencyBound is None:
highFrequencyBound = sampleRate / 2
windowing = es.Windowing(type=window,
normalized=False,
zeroPadding=zeroPadding)
spectrum = es.Spectrum()
melbands = {}
for nBands in numberBands:
melbands[nBands] = es.MelBands(
numberBands=nBands,
sampleRate=sampleRate,
lowFrequencyBound=lowFrequencyBound,
highFrequencyBound=highFrequencyBound,
inputSize=(frameSize + zeroPadding) // 2 + 1,
weighting=weighting,
normalize=normalize,
warpingFormula=warpingFormula,
type='power')
norm10k = es.UnaryOperator(type='identity', shift=1, scale=10000)
log10 = es.UnaryOperator(type='log10')
amp2db = es.UnaryOperator(type='lin2db', scale=2)
results = essentia.Pool()
for frame in es.FrameGenerator(audio,
frameSize=frameSize,
hopSize=hopSize,
startFromZero=not center):
spectrumFrame = spectrum(windowing(frame))
for nBands in numberBands:
melFrame = melbands[nBands](spectrumFrame)
results.add('mel_' + str(nBands) + '_db', amp2db(melFrame))
results.add('mel_' + str(nBands) + '_log1+10kx',
log10(norm10k(melFrame)))
return results
def FeatureExtraction_Recording(recording, params):
numBins = params.numbins
fs = params.fs
# LOAD Audio file
Audio = ess.MonoLoader(filename=recording.path, sampleRate=fs)()
Audio = ess.DCRemoval()(Audio) # PREPROCESSING / DC removal
Audio = ess.EqualLoudness()(Audio) # PREPROCESSING - Equal Loudness Filter
# Windowing Parameters (first converting from msec to number of samples)
# assuring windowSize and hopSize are even
windowSize = round(fs * params.windowSize / 1000)
windowSize = int(windowSize / 2) * 2
hopSize = round(fs * params.hopSize / 1000)
hopSize = int(hopSize / 2) * 2
tonic = float(recording.tonic)
# FRAME-BASED Spectral Analysis
hpcp = []
for frame in ess.FrameGenerator(Audio,
frameSize=windowSize,
hopSize=hopSize,
startFromZero=True):
frame = ess.Windowing(size=windowSize,
type=params.windowFunction)(frame)
mX = ess.Spectrum(size=windowSize)(frame)
mX[mX < np.finfo(float).eps] = np.finfo(float).eps
# EXTRACT frequency and magnitude information of the harmonic spectral peaks
freq, mag = ess.SpectralPeaks()(mX)
# harmonic pitch-class profiles
hpcp.append(
ess.HPCP(normalized='unitSum',
referenceFrequency=tonic,
size=numBins,
windowSize=12 / numBins)(freq, mag))
recording.chroma_framebased = np.array(hpcp)
# FEATURE SUMMARIZATION
mean_chroma = []
# global Mean of HPCP vectors
std_chroma = []
# global standard deviation of HPCP vectors
for j in range(numBins):
tmp = []
for i in range(len(recording.chroma_framebased)):
tmp.append(recording.chroma_framebased[i][j])
mean_chroma.append(np.mean(tmp))
std_chroma.append(np.std(tmp))
recording.chroma_mean = mean_chroma
recording.chroma_std = std_chroma
def extractor(filename):
fs = 44100
audio = ess.MonoLoader(filename=filename, sampleRate=fs)()
# dynamic range expansion as done in HTK implementation
audio = audio * 2**15
frameSize = 1102 # corresponds to htk default WINDOWSIZE = 250000.0
hopSize = 441 # corresponds to htk default TARGETRATE = 100000.0
fftSize = 2048
spectrumSize = fftSize // 2 + 1
zeroPadding = fftSize - frameSize
w = ess.Windowing(
type='hamming', # corresponds to htk default USEHAMMING = T
size=frameSize,
zeroPadding=zeroPadding,
normalized=False,
zeroPhase=False)
spectrum = ess.Spectrum(size=fftSize)
mfcc_htk = ess.MFCC(
inputSize=spectrumSize,
type='magnitude', # htk uses mel filterbank magniude
warpingFormula='htkMel', # htk's mel warping formula
weighting='linear', # computation of filter weights done in Hz domain
highFrequencyBound=8000, # corresponds to htk default
lowFrequencyBound=0, # corresponds to htk default
numberBands=26, # corresponds to htk default NUMCHANS = 26
numberCoefficients=13,
normalize=
'unit_max', # htk filter normaliation to have constant height = 1
dctType=3, # htk uses DCT type III
logType='log',
liftering=22) # corresponds to htk default CEPLIFTER = 22
mfccs = []
# startFromZero = True, validFrameThresholdRatio = 1 : the way htk computes windows
for frame in ess.FrameGenerator(audio,
frameSize=frameSize,
hopSize=hopSize,
startFromZero=True,
validFrameThresholdRatio=1):
spect = spectrum(w(frame))
mel_bands, mfcc_coeffs = mfcc_htk(spect)
#frame_energy = energy_func(frame)
#mfccs.append(numpy.append(mfcc_coeffs, frame_energy))
mfccs.append(mfcc_coeffs)
return mfccs
def getFeature(audio, d=True, nbf=False):
'''
MFCC of give audio interval [p[0],p[1]]
:param audio:
:param p:
:return:
'''
winAnalysis = 'hann'
# this MFCC is for pattern classification, which numberBands always be by default
MFCC40 = ess.MFCC(sampleRate=fs,
highFrequencyBound=highFrequencyBound,
inputSize=framesize + 1)
N = 2 * framesize # padding 1 time framesize
SPECTRUM = ess.Spectrum(size=N)
WINDOW = ess.Windowing(type=winAnalysis, zeroPadding=N - framesize)
mfcc = []
# audio_p = audio[p[0]*fs:p[1]*fs]
for frame in ess.FrameGenerator(audio,
frameSize=framesize,
hopSize=hopsize):
frame = WINDOW(frame)
mXFrame = SPECTRUM(frame)
bands, mfccFrame = MFCC40(mXFrame)
# mfccFrame = mfccFrame[1:]
mfcc.append(mfccFrame)
if d:
mfcc = np.array(mfcc).transpose()
dmfcc = Fdeltas(mfcc, w=5)
ddmfcc = Fdeltas(dmfcc, w=5)
feature = np.transpose(np.vstack((mfcc, dmfcc, ddmfcc)))
else:
feature = np.array(mfcc)
if not d and nbf:
mfcc = np.array(mfcc).transpose()
mfcc_out = np.array(mfcc, copy=True)
for w_r in range(1, 6):
mfcc_right_shifted = Fprev_sub(mfcc, w=w_r)
mfcc_left_shifted = Fprev_sub(mfcc, w=-w_r)
mfcc_out = np.vstack(
(mfcc_out, mfcc_left_shifted, mfcc_right_shifted))
feature = np.array(np.transpose(mfcc_out), dtype='float32')
# print feature.shape
return feature
def analyze_misc(filename, segment_duration=20):
# Compute replay gain and duration on the entire file, then load the
# segment that is centered in time with replaygain applied
audio = es.MonoLoader(filename=filename)()
replaygain = es.ReplayGain()(audio)
segment_start = (len(audio) / 44100 - segment_duration) / 2
segment_end = segment_start + segment_duration
if segment_start < 0 or segment_end > len(audio) / 44100:
raise ValueError(
'Segment duration is larger than the input audio duration')
loader = es.EasyLoader(filename=filename,
replayGain=replaygain,
startTime=segment_start,
endTime=segment_end)
windowing = es.Windowing(type='blackmanharris62')
spectrum = es.Spectrum()
powerspectrum = es.PowerSpectrum()
centroid = es.Centroid()
zcr = es.ZeroCrossingRate()
rms = es.RMS()
hfc = es.HFC()
pool = essentia.Pool()
audio = loader()
for frame in es.FrameGenerator(audio, frameSize=2048, hopSize=1024):
frame_spectrum = spectrum(windowing(frame))
pool.add('rms', rms(frame))
pool.add('rms_spectrum', rms(frame_spectrum))
pool.add('hfc', hfc(frame_spectrum))
pool.add('spectral_centroid', centroid(frame_spectrum))
pool.add('zcr', zcr(frame))
audio_st, sr, _, _, _, _ = es.AudioLoader(filename=filename)()
# Ugly hack because we don't have a StereoResample
left, right = es.StereoDemuxer()(audio_st)
resampler = es.Resample(inputSampleRate=sr, outputSampleRate=44100)
left = resampler(left)
right = resampler(right)
audio_st = es.StereoMuxer()(left, right)
audio_st = es.StereoTrimmer(startTime=segment_start,
endTime=segment_end)(audio_st)
ebu_momentary, _, _, _ = es.LoudnessEBUR128(hopSize=1024 / 44100,
startAtZero=True)(audio_st)
pool.set('ebu_momentary', ebu_momentary)
return pool
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"