def predict_onsets(self):
proc_0 = CNNOnsetProcessor()
proc_1 = OnsetPeakPickingProcessor(threshold = 0.95,fps=100)
predicts = proc_1(proc_0(self.track_name))
#=====manually adding true onsets
#predicts = [onset.prediction for onset in self.true_tablature.onsets]
#====
return list(zip(predicts, [1]*len(predicts))) # here correct it when i can get confidence
def get_onsets(path):
"""
Finds onsets of audio file
:param path: Path to audio file
:return: List of times (in seconds) corresponding to onsets
"""
processor = OnsetPeakPickingProcessor(threshold=0.45, combine=.2)
act = RNNOnsetProcessor()(path)
onsets = processor(act)
onsets = onsets.tolist()
return onsets
def GuitarOnsetDetector(audio_filename, fs=44100):
# TODO: ad hoc!!!! Derive it automatically.
onset_threshold = 2
series_delta = 0.22
fps = 180
fs = 44100
hopSize = int(fs / fps)
max_spectral_centroid = 3500
# fps must be a divisor of fs to obtain integer hopSize
# (it just simplifies the code below)
sodf = SpectralOnsetProcessor(onset_method='superflux',
fps=fps,
filterbank=LogarithmicFilterbank,
num_bands=24,
log=np.log10)
sodf_onsets = sodf(audio_filename)
# "fusion" with rms-diff.
rms, cs = rms_centroids(audio_filename,
frameSize=1024,
hopSize=hopSize,
sampleRate=fs)
rms = signal.smooth(rms, int(fs / hopSize * 0.2))
rms = preprocessing.scale(rms, with_mean=False, copy=False)
rms = rms[1:] - rms[:-1]
sodf_onsets[rms <= 0] = 0
#sodf_onsets = sodf_onsets * np.power(rms, 0.01)
#sodf_onsets[np.isnan(sodf_onsets)] = 0
proc = OnsetPeakPickingProcessor(fps=fps, threshold=onset_threshold)
p_onsets = proc(sodf_onsets)
p_onsets = combine_series(p_onsets, series_delta)
smoothed = []
for i in range(len(p_onsets)):
onset = p_onsets[i]
duration = 0.5
if (i < len(p_onsets) - 1):
duration = min((p_onsets[i + 1] - p_onsets[i]), duration)
window_len = int(duration * fs / hopSize)
s = int(float(onset) * fs / hopSize)
d = min(window_len, len(cs) - s)
w = eval('np.hanning(2*d)')
w = w[d:] / np.sum(w[d:])
w = np.reshape(w, (1, d))
c = cs[s:s + d]
smoothed.append(np.dot(w, c)[0])
result = []
for i in range(len(p_onsets)):
if smoothed[i] < max_spectral_centroid:
result.append(p_onsets[i])
return result
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)
import numpy
from madmom.features.chords import DeepChromaChordRecognitionProcessor
from madmom.audio.chroma import DeepChromaProcessor
from madmom.features.beats import DBNBeatTrackingProcessor
from madmom.features.beats import RNNBeatProcessor
from madmom.features.onsets import OnsetPeakPickingProcessor
from madmom.features.onsets import RNNOnsetProcessor
#Setting up Deep Chroma Chord Recognition Processor
dcp = DeepChromaProcessor()
decode = DeepChromaChordRecognitionProcessor()
chroma = dcp(sys.argv[1])
chords = decode(chroma)
#Setting up Onset Peak Picking Processor
proc = OnsetPeakPickingProcessor(fps=100,
threshold=0.7,
pre_avg=0.25,
post_avg=0.25,
smooth=0.01)
act = RNNOnsetProcessor()(sys.argv[1])
beats = proc(act)
#calculating msi
beatsArray = numpy.array(beats)
msi = numpy.mean(beatsArray[1:] - beatsArray[:-1]) * 1000
#generating and printing beatmap
bmaFunctions.fancyPrint(bmaFunctions.assignKeys(beats, chords, sys.argv[3]),
msi, sys.argv[2])
results_cnn[test_idx] = result.cpu().numpy()
return results_cnn
# In[ ]:
predicted = None
picked_beats = []
if PREDICT:
# beat_picker = BeatTrackingProcessor(fps=FPS) # TODO: replace with OnsetPeakPickingProcessor(fps=FPS)
beat_picker = OnsetPeakPickingProcessor(
fps=FPS,
threshold=THRESHOLD,
pre_avg=PRE_AVG,
post_avg=POST_AVG,
pre_max=PRE_MAX,
post_max=POST_MAX
) # TODO: replace with OnsetPeakPickingProcessor(fps=FPS)
# predict beats
if VERBOSE:
print('predicting...')
predicted = run_prediction(test_f) #[test_t[0], test_t[1]]
# pick peaks
if VERBOSE:
print('picking beats...')
for i, pred in enumerate(predicted):
picked = beat_picker(
def predictOneSong(audioPath, featureOption, clfModelPath):
predictions = []
if featureOption == 'convRandom':
modelSavePath = './autoencoder/savedRandomAeModels/'
features = extractRandomConvFeatures(audioPath, modelSavePath) #64 x M
elif featureOption == 'convAe':
modelSavePath = './autoencoder/savedAeModels/'
features = extractConvFeatures(audioPath, modelSavePath)
elif featureOption == 'convDae':
modelSavePath = './autoencoder/savedDaeModels/'
features = extractConvFeatures(audioPath, modelSavePath)
elif featureOption == 'baseline':
features = extractBaselineFeatures(audioPath) #60 x M
else:
print('unknown feature option')
#==== onset detection
onsetDetector = CNNOnsetProcessor()
nvt = onsetDetector(audioPath)
peakPicker = OnsetPeakPickingProcessor(fps=100)
onsets = peakPicker(nvt)
onsetsInFrames = [
round(np.divide(onset, HOPSIZE / FS)) for onset in onsets
]
#==== collect feature of interest
X = []
timeStamp = []
for i in range(0, len(onsetsInFrames)):
midIndex = int(onsetsInFrames[i])
curTime = onsets[i]
frontFrame = 0
rearFrame = 2
splicedFeature = featureSplicing(features, midIndex, frontFrame,
rearFrame)
X.append(splicedFeature)
timeStamp.append(curTime)
#print(np.shape(X))
#==== drum transcription
tmp = np.load(clfModelPath)
classifiers = tmp['arr_0']
normParams = tmp['arr_1']
clfBd = classifiers[0]
clfSd = classifiers[1]
clfHh = classifiers[2]
maxVec = normParams[0]
minVec = normParams[1]
XScaled = scaleMatrixWithMinMax(X, maxVec, minVec)
predictions = []
for i in range(0, len(timeStamp)):
curFeature = XScaled[i]
curFeature = np.expand_dims(curFeature, axis=0)
detectBd = clfBd.predict(curFeature)
detectSd = clfSd.predict(curFeature)
detectHh = clfHh.predict(curFeature)
if detectBd:
predictions.append((timeStamp[i], 'KD'))
if detectSd:
predictions.append((timeStamp[i], 'SD'))
if detectHh:
predictions.append((timeStamp[i], 'HH'))
return predictions
def rectifiedComplexDomain(filename):
audio = MonoLoader(filename=filename, sampleRate=44100)()
return OnsetPeakPickingProcessor()(SpectralOnsetProcessor(onset_method='rectified_complex_domain')(audio))
def PhaseDev(filename):
audio = MonoLoader(filename=filename, sampleRate=44100)()
return OnsetPeakPickingProcessor()(SpectralOnsetProcessor(onset_method='phase_deviation')(audio))
def modifiedKL(filename):
audio = MonoLoader(filename=filename, sampleRate=44100)()
return OnsetPeakPickingProcessor()(SpectralOnsetProcessor(onset_method='modified_kullback_leibler')(audio))
def RNNOnsetDetector(filename):
audio = MonoLoader(filename=filename, sampleRate=44100)()
return OnsetPeakPickingProcessor()(RNNOnsetProcessor()(audio))