def hfc(filename):
audio = MonoLoader(filename=filename, sampleRate=44100)()
features = []
for frame in FrameGenerator(audio, frameSize = 1024, hopSize = 512):
mag, phase =CartesianToPolar()(FFT()(Windowing(type='hann')(frame)))
features.append(OnsetDetection(method='hfc')(mag, phase))
return Onsets()(array([features]),[1])
def noveltycurve(filename):
audio = MonoLoader(filename=filename, sampleRate=44100)()
band_energy = []
for frame in FrameGenerator(audio, frameSize = 1024, hopSize = 512):
mag, phase, = CartesianToPolar()(FFT()(Windowing(type='hann')(frame)))
band_energy.append(FrequencyBands()(mag))
novelty = NoveltyCurve()(band_energy)
return Onsets()(np.array([novelty]),[1])
def calculateDownbeats(self, audio, bpm, phase):
# Step 0: calculate the CSD (Complex Spectral Difference) features
# and the associated onset detection function ON LOWPASSED SIGNAL
spec = Spectrum(size=self.FRAME_SIZE)
w = Windowing(type='hann')
fft = FFT()
c2p = CartesianToPolar()
od_csd = OnsetDetection(method='complex')
lowpass = LowPass(cutoffFrequency=1500)
pool = Pool()
# TODO test faster (numpy) way
#audio = lowpass(audio)
for frame in FrameGenerator(audio,
frameSize=self.FRAME_SIZE,
hopSize=self.HOP_SIZE):
mag, ph = c2p(fft(w(frame)))
pool.add('onsets.complex', od_csd(mag, ph))
# Step 1: normalise the data using an adaptive mean threshold
novelty_mean = self.adaptive_mean(pool['onsets.complex'], 16.0)
# Step 2: half-wave rectify the result
novelty_hwr = (pool['onsets.complex'] - novelty_mean).clip(min=0)
# Step 7 (experimental): Determine downbeat locations as subsequence with highest complex spectral difference
for i in range(4):
phase_frames = (phase * 44100.0) / (512.0)
frames = (
np.round(
np.arange(phase_frames + i * self.numFramesPerBeat(bpm),
np.size(novelty_hwr),
4 * self.numFramesPerBeat(bpm))).astype('int')
)[:
-1] # Discard last value to prevent reading beyond array (last value rounded up for example)
pool.add('output.downbeat',
np.sum(novelty_hwr[frames]) / np.size(frames))
plt.subplot(4, 1, i + 1)
plt.plot(novelty_hwr)
for f in frames:
plt.axvline(x=f)
print pool['output.downbeat']
downbeatIndex = np.argmax(pool['output.downbeat'])
plt.show()
# experimental
return 1.0 * self.beats[downbeatIndex::4]
def __call__(self, audio, SR, sumThreshold=1e-5):
self.__reset__()
if audio.ndim > 1:
audio = np.sum(audio, axis=1) / audio.ndim
fcIndexArr = []
self.hist = np.zeros(int(self.frameSize / 2 + 1))
fft = FFT(size=self.frameSize) # declare FFT function
window = Windowing(size=self.frameSize,
type="hann") # declare windowing function
self.avgFrames = np.zeros(int(self.frameSize / 2) + 1)
maxNrg = max([
sum(abs(fft(window(frame)))**2)
for frame in FrameGenerator(audio,
frameSize=self.frameSize,
hopSize=self.hopSize,
startFromZero=True)
])
for i, frame in enumerate(
FrameGenerator(audio,
frameSize=self.frameSize,
hopSize=self.hopSize,
startFromZero=True)):
frame = window(frame) # apply window to the frame
frameFft = abs(fft(frame))
nrg = sum(frameFft**2)
if nrg >= 0.1 * maxNrg:
for j in reversed(range(len(frameFft))):
if sum(frameFft[j:] / j) >= sumThreshold:
fcIndexArr.append(j)
self.hist[j] += nrg
break
self.avgFrames = self.avgFrames + frameFft
if len(fcIndexArr) == 0:
fcIndexArr.append(int(self.frameSize / 2) + 1)
self.hist[int(self.frameSize / 2)] += 1
self.avgFrames /= (i + 1)
self.mostLikelyBin, conf, binary = self.__computeMeanFc(
fcIndexArr, np.arange(int(self.frameSize / 2) + 2), hist=self.hist)
return self.mostLikelyBin * SR / self.frameSize, conf, binary
def ninos(filename,gamma=0.94):
"""
reference: Mounir, M., Karsmakers, P., & Van Waterschoot, T. (2016). Guitar note onset detection based on a spectral sparsity measure.
European Signal Processing Conference. https://doi.org/10.1109/EUSIPCO.2016.7760394
"""
N = 2048
hopSize = int(N/10)
J = int(N*gamma/2)
audio = MonoLoader(filename=filename, sampleRate=44100)()
mag = []
for frame in FrameGenerator(audio, frameSize = N, hopSize = hopSize):
m = CartesianToPolar()(FFT()(Windowing(type='hann')(frame)))[0]
m = np.asarray(m)
idx = np.argsort(m)[::-1][:J]
mag.append(m[idx])
mag = np.asarray(mag)
x2 = mag*mag
inos=np.sum(x2,axis=1)/(np.sum(x2*x2,axis=1)**(0.25))
ninos = inos/(J**(0.25))
return OnsetPeakPickingProcessor(threshold=0.03,fps=44100/hopSize)(ninos)
def detectBW(audio: list,
SR: float,
frame_size=256,
hop_size=128,
floor_db=-90,
oversample_f=1):
frame_size *= oversample_f # if an oversample factor is desired, apply it
fc_index_arr = []
fft = FFT(size=frame_size) # declare FFT function
window = Windowing(size=frame_size,
type="hann") # declare windowing function
for frame in FrameGenerator(audio,
frameSize=frame_size,
hopSize=hop_size,
startFromZero=True):
frame_fft = abs(fft(window(frame)))
frame_fft_db = 20 * np.log10(
frame_fft + eps) # calculate frame fft values in db
# compute the linear interpolation between the values of the maxima of the spectrum
interp_frame = compute_spectral_envelope(frame_fft_db, "linear")
interp_frame = modify_floor(interp_frame, floor_db, log=True)
fc_index = compute_fc(interp_frame)
if energy_verification(frame_fft, fc_index):
fc_index_arr.append(fc_index)
if len(fc_index_arr) == 0:
fc_index_arr = [frame_size]
fc_bin, conf, binary = compute_mean_fc(fc_index_arr,
np.arange(len(frame_fft)), SR)
# print("mean_fc: ", fc_bin*SR/frame_size ," conf: ", conf ," binary_result: ", binary)
return fc_bin * SR / frame_size, conf, binary
def f_essentia_extract(Audio):
## METODOS DE LIBRERIA QUE DETECTAN DONDE OCURRE CADA NOTA RESPECTO AL TIEMPO
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 = essentia.Pool()
# Computing onset detection functions.
for frame in FrameGenerator(Audio, frameSize=1024, hopSize=512):
mag, phase, = c2p(fft(w(frame)))
pool.add('features.complex', od2(mag, phase))
## inicio de cada "nota"
onsets = Onsets()
tiempos_detectados_essentia = onsets(
essentia.array([pool['features.complex']]), [1])
#print(tiempos_detectados_essentia)
return tiempos_detectados_essentia
def __init__(self):
super().__init__()
self._eq = np.array([1])
self.eq = [1]
self.fft = FFT()
self.ifft = IFFT()
from build_map import build_map sampleRate = 44100 frameSize = 2048 hopSize = 1024 windowType = "hann" mean = Mean() keyDetector = essentia.standard.Key(pcpSize=12) spectrum = Spectrum() window = Windowing(size=frameSize, zeroPadding=0, type=windowType) mfcc = MFCC() gaussian = SingleGaussian() od = OnsetDetection(method='hfc') fft = FFT() # this gives us a complex FFT c2p = CartesianToPolar() # and this turns it into a pair (magnitude, phase) onsets = Onsets(alpha=1) # dissonance spectralPeaks = SpectralPeaks(sampleRate=sampleRate, orderBy='frequency') dissonance = Dissonance() # barkbands barkbands = BarkBands(sampleRate=sampleRate) # zero crossing rate # zerocrossingrate = ZeroCrossingRate() # odd-to-even harmonic energy ratio # odd2evenharmonicenergyratio = OddToEvenHarmonicEnergyRatio()
def run(self, audio):
# TODO put this in some util class
# Step 0: calculate the CSD (Complex Spectral Difference) features
# and the associated onset detection function
spec = Spectrum(size=self.FRAME_SIZE)
w = Windowing(type='hann')
fft = FFT()
c2p = CartesianToPolar()
od_csd = OnsetDetection(method='complex')
pool = Pool()
# TODO test faster (numpy) way
for frame in FrameGenerator(audio,
frameSize=self.FRAME_SIZE,
hopSize=self.HOP_SIZE):
mag, phase = c2p(fft(w(frame)))
pool.add('onsets.complex', od_csd(mag, phase))
# Step 1: normalise the data using an adaptive mean threshold
novelty_mean = self.adaptive_mean(pool['onsets.complex'], 16.0)
# Step 2: half-wave rectify the result
novelty_hwr = (pool['onsets.complex'] - novelty_mean).clip(min=0)
# Step 3: then calculate the autocorrelation of this signal
novelty_autocorr = self.autocorr(novelty_hwr)
# Step 4: Sum over constant intervals to detect most likely BPM
valid_bpms = np.arange(self.minBpm, self.maxBpm, self.stepBpm)
for bpm in valid_bpms:
frames = (
np.round(
np.arange(0, np.size(novelty_autocorr),
self.numFramesPerBeat(bpm))).astype('int')
)[:
-1] # Discard last value to prevent reading beyond array (last value rounded up for example)
pool.add('output.bpm',
np.sum(novelty_autocorr[frames]) / np.size(frames))
bpm = valid_bpms[np.argmax(pool['output.bpm'])]
# Step 5: Calculate phase information
valid_phases = np.arange(0.0, 60.0 / bpm,
0.001) # Valid phases in SECONDS
for phase in valid_phases:
# Convert phase from seconds to frames
phase_frames = (phase * 44100.0) / (512.0)
frames = (
np.round(
np.arange(phase_frames, np.size(novelty_hwr),
self.numFramesPerBeat(bpm))).astype('int')
)[:
-1] # Discard last value to prevent reading beyond array (last value rounded up for example)
pool.add('output.phase',
np.sum(novelty_hwr[frames]) / np.size(frames))
phase = valid_phases[np.argmax(pool['output.phase'])]
print 'PHASE', phase
# Step 6: Determine the beat locations
spb = 60. / bpm #seconds per beat
beats = (np.arange(phase, (np.size(audio) / 44100) - spb + phase,
spb).astype('single'))
# Store all the results
self.bpm = bpm
self.phase = phase
self.beats = beats
self.downbeats = self.calculateDownbeats(audio, bpm, phase)