def pca(pool, namespace=''):
llspace = 'lowlevel.'
if namespace: llspace = namespace + '.lowlevel.'
sccoeffs = pool[llspace + 'sccoeffs']
scvalleys = pool[llspace + 'scvalleys']
numFrames = len(sccoeffs)
poolSc = Pool()
merged = essentia.zeros(2*len(sccoeffs[0]))
for frame in xrange(numFrames):
j = 0
for i in xrange(len(sccoeffs[frame])):
merged[j]=sccoeffs[frame][i]
merged[j+1]=scvalleys[frame][i]
j+=2
poolSc.add('contrast', merged)
poolTransformed = standard.PCA(namespaceIn='contrast',
namespaceOut='contrast')(poolSc)
contrast = poolTransformed['contrast']
pool.set(llspace+'spectral_contrast.mean', mean(contrast, axis=0))
pool.set(llspace+'spectral_contrast.var', var(contrast, axis=0))
pool.remove(llspace+'sccoeffs')
pool.remove(llspace+'scvalleys')
def get_onsets(self, _audio=[]):
if _audio != []:
audio = _audio
else:
audio = self.audio
W = es.Windowing(type=self.winType)
c2p = es.CartesianToPolar()
fft = es.FFT()
onsetDetection = es.OnsetDetection(method=self.onsetMethod,
sampleRate=44100)
onsets = es.Onsets(alpha=.2)
# onsetIndex = []
pool = Pool()
for frame in es.FrameGenerator(audio, frameSize=1024, hopSize=512):
mag, phase, = c2p(fft(W(frame)))
onsetDetection.configure(method=self.onsetMethod)
onsetFunction = onsetDetection(mag, phase)
pool.add("onsetFunction", onsetFunction)
DetectedOnsetsArray = onsets([pool["onsetFunction"]], [1])
return DetectedOnsetsArray
def pca(pool, namespace=''):
llspace = 'lowlevel.'
if namespace: llspace = namespace + '.lowlevel.'
sccoeffs = pool[llspace + 'sccoeffs']
scvalleys = pool[llspace + 'scvalleys']
numFrames = len(sccoeffs)
poolSc = Pool()
merged = essentia.zeros(2 * len(sccoeffs[0]))
for frame in xrange(numFrames):
j = 0
for i in xrange(len(sccoeffs[frame])):
merged[j] = sccoeffs[frame][i]
merged[j + 1] = scvalleys[frame][i]
j += 2
poolSc.add('contrast', merged)
poolTransformed = standard.PCA(namespaceIn='contrast',
namespaceOut='contrast')(poolSc)
contrast = poolTransformed['contrast']
pool.set(llspace + 'spectral_contrast.mean', mean(contrast, axis=0))
pool.set(llspace + 'spectral_contrast.var', var(contrast, axis=0))
pool.remove(llspace + 'sccoeffs')
pool.remove(llspace + 'scvalleys')
def get_embeddings(melspecs: dict[str, np.ndarray], architectures: dict, predictors: dict) -> Optional[dict]:
data = {}
for architecture, metadata in architectures.items():
input_pool = Pool()
input_pool.set('model/Placeholder', melspecs[metadata['essentia-algorithm']])
for dataset in metadata['datasets']:
# TODO: chunk the input melspecs to avoid OOM error
try:
output_pool = predictors[f'{dataset}-{architecture}'](input_pool)
except RuntimeError:
return None
for layer, layer_data in metadata['layers'].items():
embeddings = output_pool[layer_data['name']].squeeze()
if len(embeddings) == 0:
return None
if len(embeddings.shape) == 1:
embeddings = np.expand_dims(embeddings, axis=0)
data[f'{dataset}-{architecture}-{layer}'] = embeddings
return data
def get_onsets(self, in_filename):
# print in_filename
# Load the audio (in mono)
audio, sampleRate, numChan = AudioLoader(filename=in_filename)()
audio = MonoLoader(filename=in_filename)()
self.sampleRate = sampleRate
# 1) Compute onset detection functions
od = OnsetDetection(method='rms')
w = Windowing(type='hann')
fft = FFT()
c2p = CartesianToPolar()
pool_features = Pool()
# print 'Computing onset detection functions'
for frame in FrameGenerator(audio, frameSize=self.frame_size, hopSize=self.hop_size):
mag, phase = c2p(fft(w(frame)))
pool_features.add('features.rms', od(mag, phase))
# 2) Compute the onset locations
onsets = Onsets(silenceThreshold=0.14, delay=10)
# print 'Computing onset locations'
onsets_rms = onsets(
array([ pool_features['features.rms'] ]),
[ 1 ])
print "Num onsets: " + str(len(onsets_rms))
return onsets_rms
def compute_pitch_yin(audio):
yin = PitchYin()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('pitch_yin', yin(frame)[0])
return p['pitch_yin']
def compute_zcr(audio):
zcr = ZeroCrossingRate()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('zcr', zcr(frame))
return p['zcr']
def compute_energy(audio):
energy = Energy()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('energy', energy(frame))
return 'True'
def compute_rms(audio):
rms = RMS()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('rms', rms(frame))
return p['rms']
def compute_power_spectrum(audio):
w = Windowing(type='hann')
power_spectrum = PowerSpectrum()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('power_spectrum', power_spectrum(w(frame)))
return p['power_spectrum']
def _analyse(self, filepath):
audio = to_mono(wavread(filepath)[0])
audio = audio.astype('float32')
w = Windowing(type = 'hann')
fft = FFT() # this gives us a complex FFT
c2p = CartesianToPolar() # and this turns it into a pair (magnitude, phase)
hfc_detect = OnsetDetection(method = 'hfc')
complex_detect = OnsetDetection(method = 'complex')
rms_detect = RMS()
spec = Spectrum()
#pd = PitchDetection()
flux = Flux()
pool = Pool()
#wap = WarpedAutoCorrelation()
# let's get down to business
print 'Computing onset detection functions...'
for frame in FrameGenerator(audio, frameSize = self.frame_size,\
hopSize = self.hop_size):
mag, phase, = c2p(fft(w(frame)))
spectrum = spec(w(frame))
f = flux(spectrum)
#pitch = pd(spectrum)
pool.add('hfc', hfc_detect(mag, phase))
pool.add('complex', complex_detect(mag, phase))
pool.add('rms', rms_detect(frame))
pool.add('flux', f)
#pool.add('pitch', pitch[0])
#print pool['pitch']
#pool.add('autoc', wap(pool['pitch']))
return pool, audio
def compute_spectral_flatness(audio):
w = Windowing(type='hann')
spectrum = Spectrum()
flatness = Flatness()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('spectral_flatness', flatness(spectrum(w(frame))))
return p['spectral_flatness']
def compute_bark(audio):
w = Windowing(type='hann')
spectrum = Spectrum()
bark = BarkBands()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('bark', bark(spectrum(w(frame))))
return p['bark']
def compute_mel(audio):
w = Windowing(type='hann')
spectrum = Spectrum()
mel = MelBands(numberBands=96)
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('mel', mel(spectrum(w(frame))))
return p['mel']
def compute_pitch_yinfft(audio):
w = Windowing(type='hann')
spectrum = Spectrum()
yinfft = PitchYinFFT()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('pitch_yinfft', yinfft(spectrum(w(frame)))[0])
return p['pitch_yinfft']
def compute_spectral_rolloff(audio):
w = Windowing(type='hann')
spectrum = Spectrum()
rolloff = RollOff()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('spectral_rolloff', rolloff(spectrum(w(frame))))
return p['spectral_rolloff']
def detect_essentia(arquivo_audio,selected): #ODF using essentia library
#
try:
filename = arquivo_audio
except:
print "usage:", sys.argv[0], "<audiofile>"
sys.exit()
# don't forget, we can actually instantiate and call an algorithm on the same line!
global audio
# Phase 1: compute the onset detection function
# The OnsetDetection algorithm tells us that there are several methods available in Essentia,
# let's do two of them
if selected==3:
od = OnsetDetection(method = 'hfc')
elif selected==4:
od = OnsetDetection(method = 'complex')
elif selected==5:
od = OnsetDetection(method = 'melflux')
elif selected==6:
od = OnsetDetection(method = 'complex_phase')
elif selected==7:
od = OnsetDetection(method = 'rms')
# let's also get the other algorithms we will need, and a pool to store the results
w = Windowing(type = 'hann')
fft = FFT() # this gives us a complex FFT
c2p = CartesianToPolar() # and this turns it into a pair (magnitude, phase)
pool = Pool()
# let's get down to business
print 'Computing onset detection functions...'
for frame in FrameGenerator(audio, frameSize = 1024, hopSize = 512):
mag, phase, = c2p(fft(w(frame)))
pool.add('features.method', od(mag, phase))
# Phase 2: compute the actual onsets locations
onsets = Onsets()
print 'Computing onset times...'
onsets_method = onsets(array([ pool['features.method'] ]), [ 1 ])
# and mark them on the audio, which we'll write back to disk
# we use beeps instead of white noise to mark them, as it's more distinctive
#convertendo para o tipo list
listadet = onsets_method.tolist()
#convertendo os segundos para frames
listadet = [int(SecToFrames(x)) for x in listadet if x >= 0]
return listadet
def compute_spectral_centroid(audio):
w = Windowing(type='hann')
spectrum = Spectrum()
centroid = Centroid()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
p.add('spectral_centroid', centroid(spectrum(w(frame))))
return p['spectral_centroid']
def compute_mfcc(audio):
w = Windowing(type='hann')
spectrum = Spectrum()
mfcc = MFCC()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
_, coeffs = mfcc(spectrum(w(frame)))
p.add('mfcc', coeffs)
return p['mfcc']
def multipool(): from multiprocessing import Pool print 'Loading audio file...' audio = MonoLoader(filename = sys.argv[1])() a = AudioInfo(sys.argv[1],3,10) b = AudioInfo(sys.argv[1],4,10) c = AudioInfo(sys.argv[1],6,10) todo = [] todo.append(a) todo.append(b) todo.append(c) pool = Pool(3) pool.map(detecta,todo)
def compute_hpcp(audio):
w = Windowing(type='hann')
spectrum = Spectrum()
peaks = SpectralPeaks(orderBy='magnitude',
magnitudeThreshold=0.00001,
minFrequency=20,
maxFrequency=3500,
maxPeaks=60)
hpcp = HPCP()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=tonal_frame_size,
hopSize=tonal_hop_size,
startFromZero=True):
p.add('hpcp', hpcp(*peaks(spectrum(w(frame)))))
return p['hpcp']
def get_key(file_in):
"""
Estimates the key and scale for an audio file.
"""
loader = streaming.MonoLoader(filename=file_in)
framecutter = streaming.FrameCutter()
windowing = streaming.Windowing(type="blackmanharris62")
spectrum = streaming.Spectrum()
spectralpeaks = streaming.SpectralPeaks(orderBy="magnitude",
magnitudeThreshold=1e-05,
minFrequency=40,
maxFrequency=5000,
maxPeaks=10000)
pool = Pool()
hpcp = streaming.HPCP()
key = streaming.Key()
loader.audio >> framecutter.signal
framecutter.frame >> windowing.frame >> spectrum.frame
spectrum.spectrum >> spectralpeaks.spectrum
spectralpeaks.magnitudes >> hpcp.magnitudes
spectralpeaks.frequencies >> hpcp.frequencies
hpcp.hpcp >> key.pcp
key.key >> (pool, 'tonal.key_key')
key.scale >> (pool, 'tonal.key_scale')
key.strength >> (pool, 'tonal.key_strength')
run(loader)
return Key(pool['tonal.key_key'], pool['tonal.key_scale'])
def computeBpmHistogram(self,
noveltyCurve,
frameSize=4,
overlap=2,
frameRate=44100. / 128.,
window='hann',
zeroPadding=0,
constantTempo=False,
minBpm=30):
pool = Pool()
bpmHist = ess.BpmHistogram(frameRate=frameRate,
frameSize=frameSize,
overlap=overlap,
zeroPadding=zeroPadding,
constantTempo=constantTempo,
windowType='hann',
minBpm=minBpm)
gen = ess.VectorInput(noveltyCurve)
gen.data >> bpmHist.novelty
bpmHist.bpm >> (pool, 'bpm')
bpmHist.bpmCandidates >> (pool, 'bpmCandidates')
bpmHist.bpmMagnitudes >> (pool, 'bpmMagnitudes')
bpmHist.frameBpms >> None #(pool, 'frameBpms')
bpmHist.tempogram >> (pool, 'tempogram')
bpmHist.ticks >> (pool, 'ticks')
bpmHist.ticksMagnitude >> (pool, 'ticksMagnitude')
bpmHist.sinusoid >> (pool, 'sinusoid')
essentia.run(gen)
return pool
def get_bpm(file_in):
pool = Pool()
loader = streaming.MonoLoader(filename=file_in)
bt = streaming.RhythmExtractor2013()
bpm_histogram = streaming.BpmHistogramDescriptors()
# BPM histogram output size is 250
centroid = streaming.Centroid(range=250)
loader.audio >> bt.signal
bt.bpm >> (pool, 'bpm')
bt.ticks >> None
bt.confidence >> (pool, 'confidence')
bt.estimates >> None
bt.bpmIntervals >> bpm_histogram.bpmIntervals
bpm_histogram.firstPeakBPM >> (pool, 'bpm_first_peak')
bpm_histogram.firstPeakWeight >> None
bpm_histogram.firstPeakSpread >> None
bpm_histogram.secondPeakBPM >> (pool, 'bpm_second_peak')
bpm_histogram.secondPeakWeight >> None
bpm_histogram.secondPeakSpread >> None
bpm_histogram.histogram >> (pool, 'bpm_histogram')
bpm_histogram.histogram >> centroid.array
centroid.centroid >> (pool, 'bpm_centroid')
run(loader)
return pool['bpm']
def estimate_main_band(infile):
"""
Estimate if this is a low, mid, or high track.
Not _really_ sure if this does what I need it to,
but some quick tests looked right.
"""
loader = streaming.MonoLoader(filename=infile)
framecutter = streaming.FrameCutter()
windowing = streaming.Windowing(type="blackmanharris62")
spectrum = streaming.Spectrum()
freqbands = streaming.FrequencyBands(frequencyBands=[0, 250, 750, 4000])
pool = Pool()
loader.audio >> framecutter.signal
framecutter.frame >> windowing.frame >> spectrum.frame
spectrum.spectrum >> freqbands.spectrum
freqbands.bands >> (pool, 'bands')
run(loader)
sums = np.sum(pool['bands'], axis=0)
band = np.argmax(sums)
if band == 0:
return 'low'
elif band == 1:
return 'mid'
elif band == 2:
return 'high'
def _extract_pitch_contours(self, audio):
# Hann window with x4 zero padding
run_windowing = estd.Windowing(zeroPadding=3 * self.frame_size)
run_spectrum = estd.Spectrum(size=self.frame_size * 4)
run_spectral_peaks = estd.SpectralPeaks(
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(
binResolution=self.bin_resolution)
run_pitch_salience_function_peaks = estd.PitchSalienceFunctionPeaks(
binResolution=self.bin_resolution, minFrequency=self.min_frequency,
maxFrequency=self.max_frequency)
run_pitch_contours = estd.PitchContours(
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, 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(
pool['allframes_salience_peaks_bins'],
pool['allframes_salience_peaks_contourSaliences'])
return contours_bins, contours_start_times, contour_saliences, duration
def extractMFCCs(audio):
'''
extract mfccs from spectromra
'''
######## compute MFCCs
# maybe set highFrequencyBound=22100
frameSizeInSamples = int(round(44100 * frameSize_block))
hopSizeInSamples = int(round(44100 * hopSize_block))
inputSpectrumSize = frameSizeInSamples / 2 + 1
# inputSpectrumSize = 1025
mfcc = MFCC(numberCoefficients=num_mfccs,
numberBands=numberBands,
highFrequencyBound=highFrequencyBound,
inputSize=inputSpectrumSize)
w = Windowing(type='hann')
spectrum = Spectrum()
mfccs_array = []
pool = Pool()
audio = essentia.array(audio)
for frame in FrameGenerator(audio,
frameSize=frameSizeInSamples,
hopSize=hopSizeInSamples):
mfcc_bands, mfcc_coeffs = mfcc(spectrum(w(frame)))
pool.add('mfcc', mfcc_coeffs)
# mfccs_array = np.zeros( (len(spectogram), num_mfccs) )
# for i,spectrum in enumerate(spectogram):
#
# mfcc_bands, mfcc_coeffs = mfcc( spectrum )
# mfccs_array[i] = 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_T = essentia.array(pool['mfcc']).T
# # and plot
# imshow(mfccs_T, aspect = 'auto', interpolation='none')
# show() # unnecessary if you started "ipython --pylab"
return pool['mfcc']
def SliceDrums_BeatDetection(folder, audio_filename, fs):
od_hfc = OnsetDetection(method='hfc')
w = Windowing(type='hann')
fft = FFT() # this gives us a complex FFT
c2p = CartesianToPolar(
) # and this turns it into a pair (magnitude, phase)
onsets = Onsets()
x = MonoLoader(filename=folder + audio_filename, sampleRate=fs)()
duration = float(len(x)) / fs
x = x / np.max(np.abs(x))
t = np.arange(len(x)) / float(fs)
zero_array = t * 0 #used only for plotting purposes
#Plotting
f, axarr = plt.subplots(1, 1, figsize=(80, 20))
#Essentia beat tracking
pool = Pool()
for frame in FrameGenerator(x, frameSize=1024, hopSize=512):
mag, phase, = c2p(fft(w(frame)))
pool.add('features.hfc', od_hfc(mag, phase))
onsets_list = onsets(array([pool['features.hfc']]), [1])
axarr.vlines(onsets_list, -1, 1, color='k', zorder=2, linewidth=5.0)
axarr.plot(t, x, zorder=1)
axarr.axis('off')
for i, onset in enumerate(onsets_list):
sample = int(onset * fs) - 1000
samplename = "{}slices/{}{}__blind.wav".format(folder,
str(len(str(i))),
str(i))
if (i >= len(onsets_list) - 1):
next_sample = len(x)
else:
next_sample = int(onsets_list[i + 1] * fs) - 1000
x_seg = x[sample:next_sample]
MonoWriter(filename=samplename)(x_seg)
return onsets_list, duration
def estimate_danceability(infile):
loader = streaming.MonoLoader(filename=infile)
dance = streaming.Danceability()
pool = Pool()
loader.audio >> dance.signal
dance.danceability >> (pool, 'danceability')
run(loader)
return pool['danceability']
def get_cat_audio_pitch():
spectrum = Spectrum()
pitch = PitchYinFFT(frameSize=1024)
pool = Pool()
windowing = Windowing(type = 'hann')
cat_audio = MonoLoader(filename='cat-01.wav', sampleRate=44100)()
cat_audio_loudness = Loudness()(cat_audio)
for frame in FrameGenerator(cat_audio, frameSize=1024, hopSize=512):
spec = spectrum(windowing(frame))
p, conf = pitch(spec)
pool.add('cat_pitch', p)
cat_pitch = numpy.mean(pool['cat_pitch'])
cat_MIDI = mir_eval.multipitch.frequencies_to_midi([cat_pitch])
return cat_audio, cat_MIDI[0]
def __init__(self, arch):
self.architechture = arch
self.in_layer = None
self.out_layer = None
if arch == 'musicnn':
self.feature_extractor = es.TensorflowInputMusiCNN()
self.frame_size = 512
self.hop_size = 256
self.patch_size = 187
self.num_bands = 96
elif arch == 'vggish':
self.feature_extractor = es.TensorflowInputVGGish()
self.frame_size = 400
self.hop_size = 200
self.patch_size = 96
self.num_bands = 64
self.feature_frames = []
self.in_pool = Pool()
self.out_pool = Pool()
# setup model
self.predict = None
def compute_harmonic_magnitudes(contour_f0s, fftgram, idx_start, options):
'''
Compute for each frame harm amplitude
get harmonic partials form original spectrum
Params:
--------------------
fftgram - fftgram of whole audio file
times - ts of whole audio
hfreq - harmonics of contour
magns - magns of contour
'''
run_harm_model_anal = HarmonicModelAnal(nHarmonics=30)
# TODO: sanity check: times == len(fftgram) and contour_start_time_SAL in times
pool = Pool()
for i, contour_f0 in enumerate(contour_f0s):
if idx_start + i > len(fftgram) - 1:
sys.exit('idx start is {} while len ffmtgram is {}'.format(
idx_start, len(fftgram)))
fft = fftgram[idx_start + i]
# convert to freq :
hfreq, magn, phase = run_harm_model_anal(fft, contour_f0)
pool.add('phases', phase)
pool.add('hfreqs', hfreq)
pool.add('magns', magn)
return pool['hfreqs'], pool['magns'], pool['phases']
def harmonic_magnitudes_to_audio(hfreqs, magns, phases, options):
'''
Compute for each frame harm amplitude
convert cent bins to herz
get harmonic partials form original spectrum
Params:
hfreq - harmonics of contour
magns - magns of contour
return:
spectogram contour
out_audio_contour - audio of harmonics for a contour
'''
pool = Pool()
run_sine_model_synth = SineModelSynth(hopSize=512, sampleRate=options.Fs)
run_ifft = IFFT(size=options.windowsizeInSamples)
run_overl = OverlapAdd(frameSize=options.windowsizeInSamples,
hopSize=512,
gain=1. / options.windowsizeInSamples)
out_audio_contour = np.array(0)
for hfreq, hmag, hphase in zip(hfreqs, magns, phases):
spectrum, audio_frame = harmonics_to_audio(hfreq, hmag, hphase,
run_sine_model_synth,
run_ifft, run_overl)
out_audio_contour = np.append(out_audio_contour, audio_frame)
pool.add('spectrum', spectrum)
out_audio_contour = SM.sineModelSynth(hfreqs, magns, phases, 512, 128,
44100)
return out_audio_contour, pool['spectrum']
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 spectrogram(audio, audio_file, save_fig=True, save_fig_path=None):
if audio_file.endswith('.wav'):
w = Windowing(type='hann')
spectrum = Spectrum(
) # FFT() would return the complex FFT, here we just want the magnitude spectrum
pool = Pool()
## NOTAS -> 1 segundo de un fichero wav son aprox 90 frames y la intensidad esta dada en Hz
for frame in FrameGenerator(audio,
frameSize=1024,
hopSize=512,
startFromZero=True):
win = w(frame)
spec = spectrum(win)
pool.add('spec', spec)
aggrPool = PoolAggregator(defaultStats=['mean'])(pool)
a = sum(aggrPool['spec.mean'].T) / aggrPool['spec.mean'].T.shape[0]
# a = aggrPool['spec.mean'].T
# b = np.zeros(pool['spec'].T.shape)
b = np.array(pool['spec'].T)
# for iterator1, i in enumerate(pool['spec'].T):
# for iterator2, j in enumerate(i):
# # if j > a[iterator1]/2:
# if j > a/2 and j > 0.015:
# b[iterator1][iterator2] = j
# b = np.array([i for i in b if i.max() > 0.01])
# no para el nuevo dataset de bats
b = remove_initial_zeros(b)
b = b.tolist()
b.reverse()
b = remove_initial_zeros(b)
b.reverse()
b = np.array(b)
if save_fig:
if not save_fig_path:
save_fig_path = audio_file.replace('.wav', '_spec.jpg')
save_plots(b, save_fig_path)
return b[:200, :200].tolist()
def chromaprint(self, analysisTime=30):
"""
This algorithm computes the fingerprint of the input signal using Chromaprint algorithm.
It is a wrapper of the Chromaprint library
Returns: The chromaprints are returned as base64-encoded strings.
"""
vec_input = ess.VectorInput(self.audio_vector)
chromaprinter = ess.Chromaprinter(analysisTime=analysisTime, sampleRate=self.fs)
pool = Pool()
vec_input.data >> chromaprinter.signal
chromaprinter.fingerprint >> (pool, 'chromaprint')
run(vec_input)
return pool['chromaprint']
def compute_spectral_shape(audio):
w = Windowing(type='hann')
spectrum = Spectrum()
cm = CentralMoments()
ds = DistributionShape()
p = Pool()
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
spread, skewness, kurtosis = ds(cm(spectrum(w(frame))))
p.add('spectral_spread', spread)
p.add('spectral_skewness', skewness)
p.add('spectral_kurtosis', kurtosis)
return p['spectral_spread'], p['spectral_skewness'], p['spectral_kurtosis']
- bin 0 = 0 BPM
- bin 128 = 645.99609375 BPM
"""
sampleRate = 22050
frameSize = 8192
hopSize = 1024
rmsFrameSize = 256
rmsHopSize = 32
loader = MonoLoader(filename=input_file, sampleRate=sampleRate)
w = Windowing(type='blackmanharris62')
spectrum = Spectrum()
melbands = MelBands(sampleRate=sampleRate, numberBands=40, lowFrequencyBound=0, highFrequencyBound=sampleRate/2)
pool = Pool()
for frame in FrameGenerator(audio=loader(), frameSize=frameSize, hopSize=hopSize, startFromZero=True):
bands = melbands(spectrum(w(frame)))
pool.add('melbands', bands)
rhythmtransform = RhythmTransform(frameSize=rmsFrameSize, hopSize=rmsHopSize)
rt = rhythmtransform(pool['melbands'])
rt_mean = numpy.mean(rt, axis=0)
bin_resoluion = 5.007721656976744
print numpy.argmax(rt_mean) * bin_resoluion
if __name__ == '__main__':
opt, args = parse_args()
if len(args) != 2: #3:
print "Incorrect number of arguments\n", essentia_usage
sys.exit(1)
#profile = args[0]
input_file = args[0]
output_file = args[1]
pool = Pool()
startTime = float(opt.startTime)
endTime = float(opt.endTime)
# compute descriptors
readMetadata(input_file, pool)
INFO('Process step 1: Replay Gain')
replaygain.compute(input_file, pool, startTime, endTime)
segments_namespace=[]
if opt.segmentation:
INFO('Process step 2: Low Level')
computeLowLevel(input_file, pool, startTime, endTime)
segmentation.compute(input_file, pool, startTime, endTime)
segments = pool['segmentation.timestamps']
if __name__ == '__main__':
opt, args = parse_args()
if len(args) != 2: #3:
print "Incorrect number of arguments\n", essentia_usage
sys.exit(1)
#profile = args[0]
input_file = args[0]
output_file = args[1]
neqPool = Pool()
eqPool = Pool()
startTime = float(opt.startTime)
endTime = float(opt.endTime)
# compute descriptors
readMetadata(input_file, eqPool)
INFO('Process step 1: Replay Gain')
replaygain.compute(input_file, eqPool, startTime, endTime)
segments_namespace=[]
neqPool.merge(eqPool, 'replace')
if opt.segmentation:
INFO('Process step 2: Low Level')
computeLowLevel(input_file, neqPool, eqPool, startTime, endTime)
def run(self, fname):
citation = u"""
Atlı, H. S., Uyar, B., Şentürk, S., Bozkurt, B., and Serra, X.
(2014). Audio feature extraction for exploring Turkish makam music.
In Proceedings of 3rd International Conference on Audio Technologies
for Music and Media, Ankara, Turkey.
"""
run_windowing = Windowing(zeroPadding = 3 * self.settings.frameSize) # Hann window with x4 zero padding
run_spectrum = Spectrum(size=self.settings.frameSize * 4)
run_spectral_peaks = SpectralPeaks(minFrequency=self.settings.minFrequency,
maxFrequency = self.settings.maxFrequency,
maxPeaks = self.settings.maxPeaks,
sampleRate = self.settings.sampleRate,
magnitudeThreshold = self.settings.magnitudeThreshold,
orderBy = 'magnitude')
run_pitch_salience_function = PitchSalienceFunction(binResolution=self.settings.binResolution) # converts unit to cents, 55 Hz is taken as the default reference
run_pitch_salience_function_peaks = PitchSalienceFunctionPeaks(binResolution=self.settings.binResolution)
run_pitch_contours = PitchContours(hopSize=self.settings.hopSize,
binResolution=self.settings.binResolution,
peakDistributionThreshold = self.settings.peakDistributionThreshold)
pool = Pool();
# load audio and eqLoudness
audio = MonoLoader(filename = fname)() # MonoLoader resamples the audio signal to 44100 Hz by default
audio = EqualLoudness()(audio)
for frame in FrameGenerator(audio,frameSize=self.settings.frameSize, hopSize=self.settings.hopSize):
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_contourSaliences = run_pitch_salience_function_peaks(salience)
if not size(salience_peaks_bins):
salience_peaks_bins = array([0])
if not size(salience_peaks_contourSaliences):
salience_peaks_contourSaliences = array([0])
pool.add('allframes_salience_peaks_bins', salience_peaks_bins)
pool.add('allframes_salience_peaks_contourSaliences', salience_peaks_contourSaliences)
# post-processing: contour tracking
contours_bins, contours_contourSaliences, contours_start_times, duration = run_pitch_contours(
pool['allframes_salience_peaks_bins'],
pool['allframes_salience_peaks_contourSaliences'])
# run the simplified contour selection
#[pitch, pitch_salience] = self.ContourSelection(contours_bins,contours_contourSaliences,contours_start_times,duration)
# cent to Hz conversion
#pitch = [0. if p == 0 else 55.*(2.**(((self.settings.binResolution*(p)))/1200)) for p in pitch]
# generate time stamps
#time_stamps = [s*self.settings.hopSize/float(self.settings.sampleRate) for s in xrange(0,len(pitch))]
# [time pitch salience] matrix
#out = transpose(vstack((time_stamps, pitch, pitch_salience)))
#out = out.tolist()
# settings
settings = self.settings
settings.update({'version':self.__version__,
'slug':self.__slug__,
'source': fname,
'essentiaVersion': essentia.__version__,
'pitchUnit': 'Hz',
'citation': citation})
# matlab
#matout = cStringIO.StringIO()
#matob = {'pitch': out}
#matob.update(settings)
#scipy.io.savemat(matout, matob)
#return out
# unused
#return {'pitch': json.dumps(out),
# 'matlab': matout.getvalue(),
# 'settings': json.dumps(settings)}
return contours_bins, contours_contourSaliences, contours_start_times, duration
essentia.run(loader)
if __name__ == "__main__":
opt, args = parse_args()
if len(args) != 2: # 3:
print "Incorrect number of arguments\n", essentia_usage
sys.exit(1)
# profile = args[0]
input_file = args[0]
output_file = args[1]
neqPool = Pool()
eqPool = Pool()
startTime = float(opt.startTime)
endTime = float(opt.endTime)
# compute descriptors
readMetadata(input_file, eqPool)
INFO("Process step 1: Replay Gain")
replaygain.compute(input_file, eqPool, startTime, endTime)
segments_namespace = []
neqPool.merge(eqPool, "replace")
if opt.segmentation:
INFO("Process step 2: Low Level")
computeLowLevel(input_file, neqPool, eqPool, startTime, endTime)
def computeOnsets(inFile, outFile):
print outFile
# In this example we are going to look at how to perform some onset detection
# and mark them on the audio using the AudioOnsetsMarker algorithm.
#
# Onset detection consists of two main phases:
# 1- we need to compute an onset detection function, which is a function
# describing the evolution of some parameters, which might be representative
# of whether we might find an onset or not
# 2- performing the actual onset detection, that is given a number of these
# detection functions, decide where in the sound there actually are onsets
# we're going to work with a file specified as an argument in the command line
# try:
# filename = sys.argv[1]
# except:
# print "usage:", sys.argv[0], "<audiofile>"
# sys.exit()
# don't forget, we can actually instantiate and call an algorithm on the same line!
print "Loading audio file..."
audio = MonoLoader(filename=inFile)()
# Phase 1: compute the onset detection function
# The OnsetDetection algorithm tells us that there are several methods available in Essentia,
# let's do two of them
od1 = OnsetDetection(method="hfc")
od2 = OnsetDetection(method="complex")
# let's also get the other algorithms we will need, and a pool to store the results
w = Windowing(type="hann")
fft = FFT() # this gives us a complex FFT
c2p = CartesianToPolar() # and this turns it into a pair (magnitude, phase)
pool = Pool()
# let's get down to business
print "Computing onset detection functions..."
for frame in FrameGenerator(audio, frameSize=1024, hopSize=512):
mag, phase, = c2p(fft(w(frame)))
pool.add("features.hfc", od1(mag, phase))
pool.add("features.complex", od2(mag, phase))
# Phase 2: compute the actual onsets locations
onsets = Onsets()
print "Computing onset times..."
onsets_hfc = onsets( # this algo expects a matrix, not a vector
array([pool["features.hfc"]]),
# you need to specify weights, but as there is only a single
# function, it doesn't actually matter which weight you give it
[1],
)
np.savetxt(outFile, onsets_hfc, fmt="%f")
# Let's just take the complex as an example
onsets_complex = onsets(array([pool["features.complex"]]), [1])
np.savetxt(outFile, onsets_complex, fmt="%f")
audio = MonoLoader(filename=filename)()
# Phase 1: compute the onset detection function
# The OnsetDetection algorithm tells us that there are several methods available in Essentia,
# let's do two of them
od1 = OnsetDetection(method="hfc")
od2 = OnsetDetection(method="complex")
# let's also get the other algorithms we will need, and a pool to store the results
w = Windowing(type="hann")
fft = FFT() # this gives us a complex FFT
c2p = CartesianToPolar() # and this turns it into a pair (magnitude, phase)
pool = Pool()
# let's get down to business
print "Computing onset detection functions..."
for frame in FrameGenerator(audio, frameSize=1024, hopSize=512):
mag, phase, = c2p(fft(w(frame)))
pool.add("features.hfc", od1(mag, phase))
pool.add("features.complex", od2(mag, phase))
# Phase 2: compute the actual onsets locations
onsets = Onsets()
print "Computing onset times..."
onsets_hfc = onsets( # this algo expects a matrix, not a vector
array([pool["features.hfc"]]),
def find_files(directory, pattern):
for root, dirs, files in os.walk(directory):
for basename in files:
if basename.lower().endswith(pattern):
filename = os.path.join(root, basename)
yield filename
try:
indir = sys.argv[1]
result_file = sys.argv[2]
except:
print "usage:", sys.argv[0], "<input-directory> <result.json>"
sys.exit()
result = Pool()
files = [f for f in find_files(indir, FILE_EXT)]
print 'Found', len(files), 'audio files (' + '/'.join(FILE_EXT) + ')'
i = 0
for filename in files:
i += 1
print 'Extracting metadata:', filename
namespace = 'track_' + str(i)
try:
meta = MetadataReader(filename=filename, failOnError=True, tagPoolName=namespace + '.metadata')()
pool_meta, duration, bitrate, samplerate, channels = meta[7:]
pool_meta.set(namespace + ".file_path", os.path.relpath(filename))
pool_meta.set(namespace + ".duration", duration)
pool_meta.set(namespace + ".bit_rate", bitrate)
audio = MonoLoader(filename = filename)()
# Phase 1: compute the onset detection function
# The OnsetDetection algorithm tells us that there are several methods available in Essentia,
# let's do two of them
od1 = OnsetDetection(method = 'hfc')
od2 = OnsetDetection(method = 'complex')
# let's also get the other algorithms we will need, and a pool to store the results
w = Windowing(type = 'hann')
fft = FFT() # this gives us a complex FFT
c2p = CartesianToPolar() # and this turns it into a pair (magnitude, phase)
pool = Pool()
# let's get down to business
print 'Computing onset detection functions...'
for frame in FrameGenerator(audio, frameSize = 1024, hopSize = 512):
mag, phase, = c2p(fft(w(frame)))
pool.add('features.hfc', od1(mag, phase))
pool.add('features.complex', od2(mag, phase))
# Phase 2: compute the actual onsets locations
onsets = Onsets()
print 'Computing onset times...'
onsets_hfc = onsets(# this algo expects a matrix, not a vector
array([ pool['features.hfc'] ]),
def extractFeatures(arffDir = '.', dirname = '.', fnames = '', segment_length = 'WHOLE', hopsize = 0):
# Start to process file by file from input
for fname in fnames:
# It only process wav or mp3 file
if ".wav" not in fname.lower() and ".mp3" not in fname.lower(): continue
# Generate output dir
trackName = fname.split('/')[-1]
segmentArffDir = arffDir+"/"+trackName[:-4]+"/"
if not exists(segmentArffDir):
mkdir(segmentArffDir)
else:
print fname + ' exsits, pass...'
continue
# Read audio and some more info
loader = es.EasyLoader(filename = dirname+"/"+fname)
audio = loader.compute()
sampleRate = loader.paramValue('sampleRate')
length = int(len(audio)/sampleRate)
if length == 0: length = 1
print fname + ' length: ' + str(length)
if hopsize == 0:
hopsize = segment_length
# Specify the length of the segment
if segment_length == 'WHOLE':
step = length
end_time = length
segment_length = length
print 'The whole audio is being processed...'
else:
step = hopsize
segment_length = float(segment_length)
if step>length: continue
# Start computing segment by segment
for start_time in arange(0, length, step):
end_time = start_time + segment_length
if step != length:
print 'the time from second ' + str(start_time) + ' is being processed...'
if end_time > length:
break;
segAudio = audio[start_time*sampleRate:end_time*sampleRate]
pool = Pool()
# Setup parameters
specContrast = es.SpectralContrast(frameSize=2048, lowFrequencyBound=40, sampleRate=sampleRate)
spectrum = es.Spectrum(size=2048) #size is frameSize
mfcc = es.MFCC(lowFrequencyBound=40, sampleRate=sampleRate) # MFCC
if step > 20:
hpcp = es.HPCP(size = 12, referenceFrequency = 440, harmonics=8, bandPreset = True, minFrequency = 40.0, maxFrequency = 5000.0, \
splitFrequency = 500.0, weightType = 'cosine', nonLinear = False, windowSize = 1);# HPCP
lowLevelSpectralExtractor = \
es.LowLevelSpectralExtractor(frameSize=2048, hopSize=1024, sampleRate=sampleRate)
spectralPeaks = es.SpectralPeaks(sampleRate=sampleRate, minFrequency=40, maxFrequency=11000, maxPeaks=50, magnitudeThreshold=0.2)
# Low level spectral feature analysis
try:
features = lowLevelSpectralExtractor(segAudio)
except:
print start_time, "has failed!"
continue
# Harmonic spectral features (TODO: Is the magnitude threshold ok?)
harmonicPeaks = es.HarmonicPeaks()
pitch = es.PitchDetection() # Using YIN instead of predominant pitch analysis as this frame-based analysis
# Windowing
window = es.Windowing(size=2048)
for frame in es.FrameGenerator(segAudio, frameSize=2048, hopSize=1024):
# spectral contrast
s = spectrum(window(frame))
contrast, valley = specContrast(s)
pool.add('spectral_contrast', contrast)
pool.add('spectral_valley', valley)
# MFCC
bands, mfccs = mfcc(s)
pool.add('mfcc', mfccs[1:])
freqs, mags = spectralPeaks(s)
# HPCP
if step > 20:
hpcps = hpcp(freqs, mags)
pool.add('HPCP', hpcps)
# Self-compute spectral features
if len(freqs) > 0:
p, conf = pitch(s)
if freqs[0] == 0:
freqs = freqs[1:]
mags = mags[1:]
freqs, mags = harmonicPeaks(freqs, mags, p)
_sum = 0
if len(freqs) == 1:
specEnvelope_i = [freqs[0]] #for hsd
_sum = freqs[0]*mags[0]
elif len(freqs) == 2:
specEnvelope_i = [(freqs[0]+freqs[1])/2.0] #for hsd
_sum = freqs[0]*mags[0]+freqs[1]*mags[1]
elif len(freqs) > 2:
specEnvelope_i = [(freqs[0]+freqs[1])/2.0] #for hsd
_sum = freqs[0]*mags[0]
for i in xrange(1, len(freqs)-1):
_sum += freqs[i]*mags[i] #for hsc_i
specEnvelope_i.append((freqs[i-1]+freqs[i]+freqs[i+1])/3.0)
specEnvelope_i.append((freqs[i]+freqs[i+1])/2.0)
_sum += freqs[i+1]*mags[i+1]
hsc_i = _sum/sum(mags)
pool.add('harmonic_spectral_centroid', hsc_i)
hsd_i = sum(abs(log10(mags)-log10(specEnvelope_i)))/sum(log10(mags))
pool.add('harmonic_spectral_deviation', hsd_i)
hss_i = sqrt(sum(square(freqs-hsc_i)*square(mags))/sum(square(mags)))/hsc_i
pool.add('harmonic_spectral_spread', hss_i)
else:
pool.add('harmonic_spectral_centroid', 0)
pool.add('harmonic_spectral_deviation', 0)
pool.add('harmonic_spectral_spread', 0)
for i in xrange(0, len(features[0])):
# pool.add('barkbands', features[0][i])
pool.add('hfc', features[4][i])
pool.add('pitch', features[6][i])
pool.add('pitch_instantaneous_confidence', features[7][i])
pool.add('pitch_salience', features[8][i])
pool.add('silence_rate_20dB', features[9][i])
# pool.add('silence_rate_30dB', features[10][i])
# pool.add('silence_rate_60dB', features[11][i])
pool.add('spectral_complexity', features[12][i])
pool.add('spectral_crest', features[13][i])
pool.add('spectral_decrease', features[14][i])
pool.add('spectral_energy', features[15][i])
# pool.add('spectral_energyband_low', features[16][i])
# pool.add('spectral_energyband_middle_low', features[17][i])
# pool.add('spectral_energyband_middle_high', features[18][i])
# pool.add('spectral_energy_high', features[19][i])
pool.add('spectral_flatness_db', features[20][i])
pool.add('spectral_flux', features[21][i])
pool.add('spectral_rms', features[22][i])
pool.add('spectral_rolloff', features[23][i])
pool.add('spectral_strongpeak', features[24][i])
pool.add('zero_crossing_rate', features[25][i])
pool.add('inharmonicity', features[26][i])
pool.add('tristimulus', features[27][i])
onsetRate = es.OnsetRate()
onsets, rate = onsetRate(segAudio)
try:
aggrPool = es.PoolAggregator(defaultStats = ['mean', 'var', 'skew', 'kurt'])(pool)
except:
print start_time/step, "failed"
continue
aggrPool.add('onset_rate', rate)
#print start_time, segment_length, start_time/segment_length
fileout = segmentArffDir+trackName[:-4]+"_%003d%s"%(start_time/step, ".sig")
output = es.YamlOutput(filename = fileout)
output(aggrPool)
def run(self, fname):
citation = u'Atlı, H. S., Uyar, B., Şentürk, S., Bozkurt, B., and Serra, X. ' \
'(2014). Audio feature extraction for exploring Turkish makam music. ' \
'In Proceedings of 3rd International Conference on Audio Technologies ' \
'for Music and Media, Ankara, Turkey.'
run_windowing = Windowing(zeroPadding = 3 * self.settings.frameSize) # Hann window with x4 zero padding
run_spectrum = Spectrum(size=self.settings.frameSize * 4)
run_spectral_peaks = SpectralPeaks(minFrequency=self.settings.minFrequency,
maxFrequency = self.settings.maxFrequency,
maxPeaks = self.settings.maxPeaks,
sampleRate = self.settings.sampleRate,
magnitudeThreshold = self.settings.magnitudeThreshold,
orderBy = 'magnitude')
run_pitch_salience_function = PitchSalienceFunction(binResolution=self.settings.binResolution) # converts unit to cents, 55 Hz is taken as the default reference
run_pitch_salience_function_peaks = PitchSalienceFunctionPeaks(binResolution=self.settings.binResolution)
run_pitch_contours = PitchContours(hopSize=self.settings.hopSize,
binResolution=self.settings.binResolution,
peakDistributionThreshold = self.settings.peakDistributionThreshold)
pool = Pool();
# load audio and eqLoudness
audio = MonoLoader(filename = fname)() # MonoLoader resamples the audio signal to 44100 Hz by default
audio = EqualLoudness()(audio)
for frame in FrameGenerator(audio,frameSize=self.settings.frameSize, hopSize=self.settings.hopSize):
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_contourSaliences = run_pitch_salience_function_peaks(salience)
if not size(salience_peaks_bins):
salience_peaks_bins = array([0])
if not size(salience_peaks_contourSaliences):
salience_peaks_contourSaliences = array([0])
pool.add('allframes_salience_peaks_bins', salience_peaks_bins)
pool.add('allframes_salience_peaks_contourSaliences', salience_peaks_contourSaliences)
# post-processing: contour tracking
contours_bins, contours_contourSaliences, contours_start_times, duration = run_pitch_contours(
pool['allframes_salience_peaks_bins'],
pool['allframes_salience_peaks_contourSaliences'])
# WARNING: As of 3 April 2015, the values in "contours_start_times" leads the audio
# by 1024 + 128 samples if the read audio is in mp3 format as explained in
# https://github.com/MTG/essentia/issues/246. This roots because of the typical
# encoder/decoder problems. For now We are advancing the values in "contours_start_times"
# by 1152 samples. Uncomment the next line if this problem is fixed.
contours_start_times = [c + (1024+128)/float(self.settings.sampleRate) for c in contours_start_times]
# run the simplified contour selection
[pitch, pitch_salience] = self.ContourSelection(contours_bins,contours_contourSaliences,contours_start_times,duration)
# cent to Hz conversion
pitch = [0. if p == 0 else 55.*(2.**(((self.settings.binResolution*(p)))/1200)) for p in pitch]
# generate time stamps
time_stamps = [s*self.settings.hopSize/float(self.settings.sampleRate) for s in xrange(0,len(pitch))]
# [time pitch salience] matrix
out = transpose(vstack((time_stamps, pitch, pitch_salience)))
out = out.tolist()
# settings
settings = self.settings
settings.update({'version':self.__version__,
'slug':self.__slug__,
'source': fname,
'essentiaVersion': essentia.__version__,
'pitchUnit': 'Hz',
'citation': citation})
# matlab
matout = cStringIO.StringIO()
matob = {'pitch': out}
matob.update(settings)
scipy.io.savemat(matout, matob)
return {'pitch': json.dumps(out),
'matlab': matout.getvalue(),
'settings': json.dumps(settings)}
def deteccoes(arquivo_audio): #Return a list with all detections
try:
filename = sys.argv[1]
except:
print "usage:", sys.argv[0], "<audiofile>"
sys.exit()
audio = MonoLoader(filename = filename)()
# Phase 1: compute the onset detection function
# The OnsetDetection algorithm tells us that there are several methods available in Essentia,
# let's do two of them
od1 = OnsetDetection(method = 'hfc')
od2 = OnsetDetection(method = 'complex')
# let's also get the other algorithms we will need, and a pool to store the results
w = Windowing(type = 'hann')
fft = FFT() # this gives us a complex FFT
c2p = CartesianToPolar() # and this turns it into a pair (magnitude, phase)
pool = Pool()
# let's get down to business
print 'Computing onset detection functions...'
for frame in FrameGenerator(audio, frameSize = 1024, hopSize = 512):
mag, phase, = c2p(fft(w(frame)))
pool.add('features.hfc', od1(mag, phase))
pool.add('features.complex', od2(mag, phase))
# Phase 2: compute the actual onsets locations
onsets = Onsets()
print 'Computing onset times...'
onsets_hfc = onsets(# this algo expects a matrix, not a vector
array([ pool['features.hfc'] ]),
# you need to specify weights, but as there is only a single
# function, it doesn't actually matter which weight you give it
[ 1 ])
onsets_complex = onsets(array([ pool['features.complex'] ]), [ 1 ])
# and mark them on the audio, which we'll write back to disk
# we use beeps instead of white noise to mark them, as it's more distinctive
print 'Writing audio files to disk with onsets marked...'
# mark the 'hfc' onsets:
#convertendo para o tipo list
listadethfc = onsets_hfc.tolist()
listadetcomplex = onsets_complex.tolist()
#convertendo os segundos para frames
listadethfc = [int(SecToFrames(x)) for x in listadethfc if x >= 0]
listadetcomplex = [int(SecToFrames(x)) for x in listadetcomplex if x >= 0]
return listadetcomplex
