def energyThresholdAudio(soundfilesList):
for sound in soundfilesList:
RMS = esst.RMS()
audioLoader = esst.MonoLoader(filename=sound)
audio = audioLoader()
start=0
end=0
thresh=0.05
rms_vals=[]
for frame in esst.FrameGenerator(audio, frameSize=2048, hopSize=1024, startFromZero=True):
rms = RMS(frame)
rms_vals.append(float(rms))
rms_vals = np.array(rms_vals)
higher=np.where(rms_vals >= thresh)[0]
if len(higher) > 1:
start=higher[0]
end=higher[-1]
else:
continue
newAudio = audio[start*1024:end*1024]
writer = esst.MonoWriter(filename=sound, format="mp3")
writer(newAudio)
print (sound)
def rms(audio,params):
""" hop size, frame size, window type """
hopSize, frameSize, wtype = params
w = Windowing(type=wtype)
spec = Spectrum()
result = []
RMS = ess.RMS()
for frame in ess.FrameGenerator(audio, frameSize = frameSize, hopSize = hopSize):
sf = spec(w(frame))
result.append(RMS(sf))
return np.asarray(result),hopSize
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 postProcessTicks(audioFilename, ticks, ticksAmp, pool):
'''Computes delta energy in order to find the correct position of the ticks'''
# get rid of beats of beats > audio.length
# if t < 0 or t > pool['length']: continue
# ticks.append(float(t))
# ticksAmp.append(float(amp))
#ticks = essentia.postProcessTicks(ticks, ticksAmp, 60./pool['harmonicBpm'][0]);
beatWindowDuration = 0.01 # seconds
beatDuration = 0.005 # seconds
rmsFrameSize = 64
rmsHopSize = rmsFrameSize / 2
audio = std.MonoLoader(filename=audioFilename,
sampleRate=pool['samplerate'],
downmix=pool['downmix'])()
for i, tick in enumerate(ticks):
startTime = tick - beatWindowDuration / 2.0
if startTime < 0: startTime = 0
endTime = startTime + beatWindowDuration + beatDuration + 0.0001
slice = std.Trimmer(sampleRate=pool['samplerate'],
startTime=startTime,
endTime=endTime)(audio)
frames = std.FrameGenerator(slice,
frameSize=rmsFrameSize,
hopSize=rmsHopSize)
maxDeltaRms = 0
RMS = std.RMS()
prevRms = 0
pos = 0
tickPos = pos
for frame in frames:
rms = RMS(frame)
diff = rms - prevRms
if diff > maxDeltaRms:
tickPos = pos
maxDeltaRms = diff
pos += 1
prevRms = rms
ticks[i] = tick + tickPos * float(rmsHopSize) / pool['samplerate']
return ticks
def highEnergyFrames(audioIn, threshold=0.05, strip=False):
# strip: if True, only low-energy frames at the beginning and at the end are discarded
RMS = ess.RMS()
highEnergyAudio = []
frames = ess.FrameGenerator(audioIn,
frameSize=2048,
hopSize=1024,
startFromZero=True)
rmsValues = np.array([float(RMS(frame)) for frame in frames])
highRMSFrames = np.where(rmsValues > threshold)[0]
outSamples = [frame * 1024 for frame in highRMSFrames]
if strip:
# return the middle section of the audio
highEnergyAudio = audioIn[outSamples[0]:outSamples[-1]] if len(
highRMSFrames > 1) else audioIn
else:
# return all high-energy samples
highEnergyAudio = audioIn[outSamples]
return highEnergyAudio
## Tabla analysis and synthesis module for the HAMR 2015 ISMIR hack
import os
import essentia as es
import essentia.standard as ess
import numpy as np
import pickle
import glob
import utilFunctions as UF
import scipy.spatial.distance as DS
import parameters as params
import csv
rms=ess.RMS()
window = ess.Windowing(type = "hamming")
spec = ess.Spectrum(size=params.Nfft)
zz = np.zeros((params.zeropadLen,), dtype = 'float32')
genmfcc = ess.MFCC(highFrequencyBound = 22000.0, inputSize = params.Nfft/2+1, sampleRate = params.Fs)
hps = ess.HighPass(cutoffFrequency = 240.0)
onsets = ess.Onsets()
strokeLabels = ['dha', 'dhen', 'dhi', 'dun', 'ge', 'kat', 'ke', 'na', 'ne', 're', 'tak', 'te', 'tit', 'tun']
taals = {"teen": {"nmatra": 16, "accents": np.array([4, 1, 1, 1, 3, 1, 1, 1, 2, 1, 1, 1, 3, 1, 1, 1])},
"ek": {"nmatra": 12, "accents": np.array([4, 1, 1, 2, 1, 1, 3, 1, 1, 2, 1, 1])},
"jhap": {"nmatra": 10, "accents": np.array([4, 1, 2, 1, 1, 3, 1, 2, 1, 1])},
"rupak": {"nmatra": 7, "accents": np.array([2, 1, 1, 3, 1, 3, 1])}
}
rolls = [{"bol": ['dha/dha_02', 'te/te_05', 're/re_04', 'dha/dha_02'], "dur": np.array([1.0, 1.0, 1, 1]), "amp": np.array([1.0, 1.0, 1.0, 1.0])},
def compute(audio, pool, options):
# analysis parameters
sampleRate = options['sampleRate']
frameSize = options['frameSize']
hopSize = options['hopSize']
windowType = options['windowType']
# temporal descriptors
lpc = ess.LPC(order=10, type='warped', sampleRate=sampleRate)
zerocrossingrate = ess.ZeroCrossingRate()
# frame algorithms
frames = ess.FrameGenerator(audio=audio, frameSize=frameSize, hopSize=hopSize)
window = ess.Windowing(size=frameSize, zeroPadding=0, type=windowType)
spectrum = ess.Spectrum(size=frameSize)
# spectral algorithms
barkbands = ess.BarkBands(sampleRate=sampleRate)
centralmoments = ess.CentralMoments()
crest = ess.Crest()
centroid = ess.Centroid()
decrease = ess.Decrease()
spectral_contrast = ess.SpectralContrast(frameSize=frameSize,
sampleRate=sampleRate,
numberBands=6,
lowFrequencyBound=20,
highFrequencyBound=11000,
neighbourRatio=0.4,
staticDistribution=0.15)
distributionshape = ess.DistributionShape()
energy = ess.Energy()
# energyband_bass, energyband_middle and energyband_high parameters come from "standard" hi-fi equalizers
energyband_bass = ess.EnergyBand(startCutoffFrequency=20.0, stopCutoffFrequency=150.0, sampleRate=sampleRate)
energyband_middle_low = ess.EnergyBand(startCutoffFrequency=150.0, stopCutoffFrequency=800.0, sampleRate=sampleRate)
energyband_middle_high = ess.EnergyBand(startCutoffFrequency=800.0, stopCutoffFrequency=4000.0,
sampleRate=sampleRate)
energyband_high = ess.EnergyBand(startCutoffFrequency=4000.0, stopCutoffFrequency=20000.0, sampleRate=sampleRate)
flatnessdb = ess.FlatnessDB()
flux = ess.Flux()
harmonic_peaks = ess.HarmonicPeaks()
hfc = ess.HFC()
mfcc = ess.MFCC()
rolloff = ess.RollOff()
rms = ess.RMS()
strongpeak = ess.StrongPeak()
# pitch algorithms
pitch_detection = ess.PitchYinFFT(frameSize=frameSize, sampleRate=sampleRate)
pitch_salience = ess.PitchSalience()
# dissonance
spectral_peaks = ess.SpectralPeaks(sampleRate=sampleRate, orderBy='frequency')
dissonance = ess.Dissonance()
# spectral complexity
# magnitudeThreshold = 0.005 is hardcoded for a "blackmanharris62" frame
spectral_complexity = ess.SpectralComplexity(magnitudeThreshold=0.005)
INFO('Computing Low-Level descriptors...')
# used for a nice progress display
total_frames = frames.num_frames()
n_frames = 0
start_of_frame = -frameSize * 0.5
pitches, pitch_confidences = [], []
progress = Progress(total=total_frames)
#scPool = es.Pool() # pool for spectral contrast
for frame in frames:
frameScope = [start_of_frame / sampleRate, (start_of_frame + frameSize) / sampleRate]
# pool.setCurrentScope(frameScope)
# silence rate
# pool.add(namespace + '.' + 'silence_rate_60dB', es.isSilent(frame))
pool.add(namespace + '.' + 'silence_rate_60dB', is_silent_threshold(frame, -60))
pool.add(namespace + '.' + 'silence_rate_30dB', is_silent_threshold(frame, -30))
pool.add(namespace + '.' + 'silence_rate_20dB', is_silent_threshold(frame, -20))
if options['skipSilence'] and es.isSilent(frame):
total_frames -= 1
start_of_frame += hopSize
continue
# temporal descriptors
pool.add(namespace + '.' + 'zerocrossingrate', zerocrossingrate(frame))
(frame_lpc, frame_lpc_reflection) = lpc(frame)
pool.add(namespace + '.' + 'temporal_lpc', frame_lpc)
frame_windowed = window(frame)
frame_spectrum = spectrum(frame_windowed)
# spectrum-based descriptors
power_spectrum = frame_spectrum ** 2
pool.add(namespace + '.' + 'spectral_centroid', centroid(power_spectrum))
pool.add(namespace + '.' + 'spectral_decrease', decrease(power_spectrum))
pool.add(namespace + '.' + 'spectral_energy', energy(frame_spectrum))
pool.add(namespace + '.' + 'spectral_energyband_low', energyband_bass(frame_spectrum))
pool.add(namespace + '.' + 'spectral_energyband_middle_low', energyband_middle_low(frame_spectrum))
pool.add(namespace + '.' + 'spectral_energyband_middle_high', energyband_middle_high(frame_spectrum))
pool.add(namespace + '.' + 'spectral_energyband_high', energyband_high(frame_spectrum))
pool.add(namespace + '.' + 'hfc', hfc(frame_spectrum))
pool.add(namespace + '.' + 'spectral_rms', rms(frame_spectrum))
pool.add(namespace + '.' + 'spectral_flux', flux(frame_spectrum))
pool.add(namespace + '.' + 'spectral_rolloff', rolloff(frame_spectrum))
pool.add(namespace + '.' + 'spectral_strongpeak', strongpeak(frame_spectrum))
# central moments descriptors
frame_centralmoments = centralmoments(power_spectrum)
(frame_spread, frame_skewness, frame_kurtosis) = distributionshape(frame_centralmoments)
pool.add(namespace + '.' + 'spectral_kurtosis', frame_kurtosis)
pool.add(namespace + '.' + 'spectral_spread', frame_spread)
pool.add(namespace + '.' + 'spectral_skewness', frame_skewness)
# dissonance
(frame_frequencies, frame_magnitudes) = spectral_peaks(frame_spectrum)
frame_dissonance = dissonance(frame_frequencies, frame_magnitudes)
pool.add(namespace + '.' + 'dissonance', frame_dissonance)
# mfcc
(frame_melbands, frame_mfcc) = mfcc(frame_spectrum)
pool.add(namespace + '.' + 'mfcc', frame_mfcc)
# spectral contrast
(sc_coeffs, sc_valleys) = spectral_contrast(frame_spectrum)
#scPool.add(namespace + '.' + 'sccoeffs', sc_coeffs)
#scPool.add(namespace + '.' + 'scvalleys', sc_valleys)
pool.add(namespace + '.' + 'spectral_contrast', sc_coeffs)
# barkbands-based descriptors
frame_barkbands = barkbands(frame_spectrum)
pool.add(namespace + '.' + 'barkbands', frame_barkbands)
pool.add(namespace + '.' + 'spectral_crest', crest(frame_barkbands))
pool.add(namespace + '.' + 'spectral_flatness_db', flatnessdb(frame_barkbands))
barkbands_centralmoments = ess.CentralMoments(range=len(frame_barkbands) - 1)
(barkbands_spread, barkbands_skewness, barkbands_kurtosis) = distributionshape(
barkbands_centralmoments(frame_barkbands))
pool.add(namespace + '.' + 'barkbands_spread', barkbands_spread)
pool.add(namespace + '.' + 'barkbands_skewness', barkbands_skewness)
pool.add(namespace + '.' + 'barkbands_kurtosis', barkbands_kurtosis)
# pitch descriptors
frame_pitch, frame_pitch_confidence = pitch_detection(frame_spectrum)
if frame_pitch > 0 and frame_pitch <= 20000.:
pool.add(namespace + '.' + 'pitch', frame_pitch)
pitches.append(frame_pitch)
pitch_confidences.append(frame_pitch_confidence)
pool.add(namespace + '.' + 'pitch_instantaneous_confidence', frame_pitch_confidence)
frame_pitch_salience = pitch_salience(frame_spectrum[:-1])
pool.add(namespace + '.' + 'pitch_salience', frame_pitch_salience)
# spectral complexity
pool.add(namespace + '.' + 'spectral_complexity', spectral_complexity(frame_spectrum))
# display of progress report
progress.update(n_frames)
n_frames += 1
start_of_frame += hopSize
# if no 'temporal_zerocrossingrate' it means that this is a silent file
if 'zerocrossingrate' not in descriptorNames(pool.descriptorNames(), namespace):
raise ess.EssentiaError('This is a silent file!')
#spectralContrastPCA(scPool, pool)
# build pitch value histogram
from math import log
from numpy import bincount
# convert from Hz to midi notes
midipitches = []
unknown = 0
for freq in pitches:
if freq > 0. and freq <= 12600:
midipitches.append(12 * (log(freq / 6.875) / 0.69314718055995) - 3.)
else:
unknown += 1
if len(midipitches) > 0:
# compute histogram
midipitchhist = bincount(midipitches)
# set 0 midi pitch to be the number of pruned value
midipitchhist[0] = unknown
# normalise
midipitchhist = [val / float(sum(midipitchhist)) for val in midipitchhist]
# zero pad
for i in range(128 - len(midipitchhist)): midipitchhist.append(0.0)
else:
midipitchhist = [0.] * 128
midipitchhist[0] = 1.
# pitchhist = ess.array(zip(range(len(midipitchhist)), midipitchhist))
pool.add(namespace + '.' + 'spectral_pitch_histogram', midipitchhist) # , pool.GlobalScope)
# the code below is the same as the one above:
# for note in midipitchhist:
# pool.add(namespace + '.' + 'spectral_pitch_histogram_values', note)
# print "midi note:", note
pitch_centralmoments = ess.CentralMoments(range=len(midipitchhist) - 1)
(pitch_histogram_spread, pitch_histogram_skewness, pitch_histogram_kurtosis) = distributionshape(
pitch_centralmoments(midipitchhist))
pool.add(namespace + '.' + 'spectral_pitch_histogram_spread', pitch_histogram_spread) # , pool.GlobalScope)
progress.finish()
def rms(self, audio):
rmsFunc = es.RMS()
return 20 * math.log10(rmsFunc(audio))
import matplotlib.pyplot as plt
mbid = 'ead85d20-ce7d-4ed0-a00d-0ae199b94d12'
hopSize = 128
loader = es.MonoLoader(filename=os.path.join(mbid, mbid + '-voice.mp3'))
track = loader()
# track[track<0.000001] = 0
print(len(track) / 44100.)
# pY = es.PitchYin(minFrequency=55, maxFrequency=900, tolerance=0.06)
pY = es.PitchYin(minFrequency=55, maxFrequency=600, tolerance=0.03)
rms = es.RMS()
pitch = []
loudness = []
print('Computing pitch and loudness')
for frame in es.FrameGenerator(track,
frameSize=2048,
hopSize=hopSize,
startFromZero=True):
f = pY(frame)
if f[1] >= 0.8:
pitch.append(f[0])
loudness.append(rms(frame))
else:
pitch.append(0)
def compute_features(complete_path):
result = []
meta_result = []
file_count = 0
# for loop over files
for file in os.listdir(complete_path):
if file.endswith(".wav"):
file_count+=1
# print(file +' : ' + str(file_count))
# load our audio into an array
audio = es.MonoLoader(filename=complete_path + file, sampleRate=44100)()
# create the pool and the necessary algorithms
pool = essentia.Pool()
window = es.Windowing()
energy = es.Energy()
spectrum = es.Spectrum()
centroid = es.Centroid(range=22050)
rolloff = es.RollOff()
crest = es.Crest()
speak = es.StrongPeak()
rmse = es.RMS()
mfcc = es.MFCC()
flux = es.Flux()
barkbands = es.BarkBands( sampleRate = 44100)
zerocrossingrate = es.ZeroCrossingRate()
meta = es.MetadataReader(filename=complete_path + file, failOnError=True)()
pool_meta, duration, bitrate, samplerate, channels = meta[7:]
# centralmoments = es.SpectralCentralMoments()
# distributionshape = es.DistributionShape()
# compute the centroid for all frames in our audio and add it to the pool
for frame in es.FrameGenerator(audio, frameSize = 1024, hopSize = 512):
frame_windowed = window(frame)
frame_spectrum = spectrum(frame_windowed)
c = centroid(frame_spectrum)
pool.add('spectral.centroid', c)
cr = crest(frame_spectrum)
pool.add('spectral crest', cr)
r = rolloff(frame_spectrum)
pool.add('spectral rolloff', r)
sp = speak(frame_spectrum)
pool.add('strong peak', sp)
rms = rmse(frame_spectrum)
pool.add('RMS', rms)
pool.add('spectral_energy', energy(frame_spectrum))
# (frame_melbands, frame_mfcc) = mfcc(frame_spectrum)
# pool.add('frame_MFCC', frame_mfcc)
fl = flux(frame_spectrum)
pool.add('spectral flux', fl)
# bbands = barkbands(frame_spectrum)
# pool.add('bark bands', bbands)
zcr = zerocrossingrate(frame_spectrum)
pool.add('zero crossing rate', zcr)
# frame_centralmoments = centralmoments(power_spectrum)
# (frame_spread, frame_skewness, frame_kurtosis) = distributionshape(frame_centralmoments)
# pool.add('spectral_kurtosis', frame_kurtosis)
# pool.add('spectral_spread', frame_spread)
# pool.add('spectral_skewness', frame_skewness)
# aggregate the results (find mean if needed)
aggrpool = es.PoolAggregator(defaultStats = ['mean'])(pool) #,'stdev' ])(pool)
pool_meta.set("duration", duration)
pool_meta.set("filename", os.path.relpath(file))
# write pools to lists
pool_arr = pool_to_array(aggrpool)
result.append(pool_arr)
meta_arr = pool_to_array(pool_meta)
meta_result.append(meta_arr)
features_df = pd.DataFrame.from_records(result)
features_df.columns = ['centroid', 'crest','roll off','strong peak','rms','energy','flux','zcr']
meta_df = pd.DataFrame.from_records(meta_result)
meta_df.columns = ['duration','filename','metadata.tags.comment']
del meta_df['metadata.tags.comment']
return features_df,meta_df
