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 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 __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,
signal,
sampleRate,
frameSize=1024,
hopSize=512,
method='complex',
window='hann'):
self.signal = signal.astype(np.float32)
self.sampleRate = sampleRate
self.frameSize = frameSize
self.hopSize = hopSize
self.calcOnsetFunc = es.OnsetDetection(method=method)
self.window = es.Windowing(type=window)
def OnsetsSegmentation(audio,
frame_size=1024,
frame_hop=512,
windowing_type='hann',
onsets_method='hfc'):
#declaração dos algoritmos que serão usados
spec = es_mode.Spectrum()
fft = es_mode.FFT()
c2p = es_mode.CartesianToPolar()
od1 = es_mode.OnsetDetection(method=onsets_method)
w = es_mode.Windowing(type=windowing_type)
pool = es.Pool()
#Função que será executada a cada frame
def F(n):
spectrum = spec(w(n))
mag, phase, = c2p(fft(w(n)))
pool.add('features.spectrum', spectrum)
pool.add('features.', phase)
pool.add('features.onsetdetection', od1(spectrum, phase))
#define a função contínua de onsets para cada frame
qtdFrames = inFrames(audio=audio,
algorithm=F,
frameSize=frame_size,
hopSize=frame_hop)
#print("Quantidade de frames: ", qtdFrames)
audio_duration = es_mode.Duration()(audio)
frame_rate = qtdFrames / audio_duration
os = es_mode.Onsets(frameRate=frame_rate)
#matriz de algoritmos de detecção de onset executados
onset_detection_matrix = es.array([pool['features.onsetdetection']])
#segundo parâmetro é o vetor de pesos para cada detecção de onset
onsets = os(onset_detection_matrix, [1])
end_times = es.array(np.append(onsets, audio_duration))
start_times = es.array(np.append([0], onsets))
segments = es_mode.Slicer(endTimes=end_times,
startTimes=start_times,
timeUnits="seconds")(audio)
return segments, onsets
def __detect_onsets(self, file, frame_size, hop_size, windowfnc,
normalize) -> None:
window = estd.Windowing(size=frame_size,
type=windowfnc.value,
normalized=normalize)
fft = estd.FFT(size=frame_size)
pool = es.Pool()
pool_add = pool.add
cart_to_polar = estd.CartesianToPolar()
detect_onset = estd.OnsetDetection(method=self.algo)
for frame in estd.FrameGenerator(file.audio,
frameSize=frame_size,
hopSize=hop_size):
mag, phase, = cart_to_polar(fft(window(frame)))
pool_add(
"features." + self.algo,
detect_onset(mag, phase),
)
# The onsets algo expects a matrix of features which can be weighted
self.onsets = estd.Onsets()(es.array([pool["features." + self.algo]]),
[1])
def detect_onset(audio, index):
# should be able to fetch the module from cache
import essentia.standard as ess_std
from essentia import array
print("Subprocess {} starts".format(index))
processing_start = time()
onset_detector = ess_std.OnsetDetection(method="complex")
window = ess_std.Windowing(type="hann")
fft = ess_std.FFT()
c2p = ess_std.CartesianToPolar()
onsets = ess_std.Onsets()
frames = []
for frame in ess_std.FrameGenerator(audio, frameSize=1024, hopSize=512):
mag, phase = c2p(fft(window(frame)))
frames.append(onset_detector(mag, phase))
onsets_array = onsets(array([frames]), [1])
print("Subprocess {} finished. Elapsed time: {:.2}s".format(
index,
time() - processing_start))
return onsets_array
import numpy as np
import matplotlib.pyplot as plt
import essentia.standard as ess
M = 1024
N = 1024
H = 512
fs = 44100
spectrum = ess.Spectrum(size=N)
window = ess.Windowing(size=M, type='hann')
flux = ess.Flux()
onsetDetection = ess.OnsetDetection(method='hfc')
x = ess.MonoLoader(filename='../../../sounds/speech-male.wav', sampleRate=fs)()
fluxes = []
onsetDetections = []
for frame in ess.FrameGenerator(x, frameSize=M, hopSize=H, startFromZero=True):
mX = spectrum(window(frame))
flux_val = flux(mX)
fluxes.append(flux_val)
onsetDetection_val = onsetDetection(mX, mX)
onsetDetections.append(onsetDetection_val)
onsetDetections = np.array(onsetDetections)
fluxes = np.array(fluxes)
plt.figure(1, figsize=(9.5, 7))
plt.subplot(2, 1, 1)
plt.plot(np.arange(x.size) / float(fs), x)
plt.axis([0, x.size / float(fs), min(x), max(x)])
plt.ylabel('amplitude')
def __init__(self, params, fsm=None):
self.onset_threshold = params['onset_threshold']
self.offset_threshold = params['offset_threshold']
self.max_attack_time = params['max_attack_time']
self.max_release_time = params['max_release_time']
self.attack_slope_ratio = params['attack_slope_ratio']
self.release_slope_ratio = params['release_slope_ratio']
self.flux_threshold = params['flux_threshold']
self.mel_threshold = params['mel_threshold']
self.rms_threshold = params['rms_threshold']
self.conf_threshold = params['conf_threshold']
self.ratio_mel = params['ratio_mel']
self.ratio_rms = params['ratio_rms']
self.rms_threshold_value = 0
self.mel_threshold_vale = 0
self.fs = params['fs']
self.hop_size = params['hop_size']
self.max_attack_frames = seconds2frames(self.max_attack_time,
fs=self.fs,
hop_size=self.hop_size)
self.max_release_frames = seconds2frames(self.max_release_time,
fs=self.fs,
hop_size=self.hop_size)
self.ext_fsm = fsm # external state machine to send events to
self.buffer = []
self.was_onset = False
self.was_offset = False
self.onset_counter = self.offset_counter = None
self.onset_samples = 2 # number of consecutive samples to be above threshold
self.offset_samples = 3 # number of consecutive samples to be below threshold
self.peak_detect = GrowingSlopeEnd(max_frames=self.max_attack_frames,
m=self.attack_slope_ratio)
# essentia algorithms initialization
self.o_mel = estd.OnsetDetection(method='melflux')
self.o_rms = estd.OnsetDetection(method='rms')
self.o_hfc = estd.OnsetDetection(method='hfc')
self.o_flux = estd.OnsetDetection(method='flux')
self.o_complex = estd.OnsetDetection(method='complex')
self.fft = estd.FFT()
self.c2p = estd.CartesianToPolar()
self.w = estd.Windowing(type='hann')
# STATE MACHINE
self.fsm = Fysom({
'initial':
'detecting',
'events': [{
'name': 'onset',
'src': 'detecting',
'dst': 'attack'
}, {
'name': 'peak',
'src': 'attack',
'dst': 'sustain'
}, {
'name': 'offset',
'src': 'sustain',
'dst': 'detecting'
}, {
'name': 'reset',
'src': ['detecting', 'attack', 'sustain'],
'dst': 'detecting'
}],
'callbacks': {
'ondetecting': self.on_detecting,
'onattack': self.on_attack,
'onsustain': self.on_sustain,
'onbeforeonset': self.on_onset,
'onbeforepeak': self.on_peak,
'onbeforeoffset': self.on_offset
}
})
