class MonoNote(object):
def __init__(self):
self.hmm = MonoNoteHMM()
def process(self, pitchProb):
obsProb = [
self.hmm.calculatedObsProb(pitchProb[0]),
]
for iFrame in range(1, len(pitchProb)):
obsProb += [self.hmm.calculatedObsProb(pitchProb[iFrame])]
out = []
path, scale = self.hmm.decodeViterbi(obsProb)
for iFrame in range(len(path)):
currPitch = -1.0
stateKind = 0
currPitch = self.hmm.par.minPitch + (
path[iFrame] / self.hmm.par.nSPP) * 1.0 / self.hmm.par.nPPS
stateKind = (path[iFrame]) % self.hmm.par.nSPP + 1
out.append(FrameOutput(iFrame, currPitch, stateKind))
return out
def __init__(self, STEPS_PER_SEMITONE, NUM_SEMITONES,
with_bar_dependent_probs, hopTime, usul_type):
'''
create hmm and build its transition matrix
'''
self.with_bar_dependent_probs = with_bar_dependent_probs
self.hmm = MonoNoteHMM(STEPS_PER_SEMITONE, NUM_SEMITONES,
with_bar_dependent_probs, hopTime, usul_type)
self.hmm.build_trans_probs(with_bar_dependent_probs)
self.hmm.build_obs_model()
class MonoNote(object):
def __init__(self):
self.hmm = MonoNoteHMM()
def process(self, pitchProb):
obsProb = [self.hmm.calculatedObsProb(pitchProb[0]), ]
for iFrame in range(1, len(pitchProb)):
obsProb += [self.hmm.calculatedObsProb(pitchProb[iFrame])]
out = []
path, scale = self.hmm.decodeViterbi(obsProb)
for iFrame in range(len(path)):
currPitch = -1.0
stateKind = 0
currPitch = self.hmm.par.minPitch + (path[iFrame]/self.hmm.par.nSPP) * 1.0/self.hmm.par.nPPS
stateKind = (path[iFrame]) % self.hmm.par.nSPP + 1
out.append(FrameOutput(iFrame, currPitch, stateKind))
return out
def __init__(self):
self.hmm = MonoNoteHMM()
def __init__(self):
self.hmm = MonoNoteHMM()
class MonoNote(object):
def __init__(self, STEPS_PER_SEMITONE, NUM_SEMITONES,
with_bar_dependent_probs, hopTime, usul_type):
'''
create hmm and build its transition matrix
'''
self.with_bar_dependent_probs = with_bar_dependent_probs
self.hmm = MonoNoteHMM(STEPS_PER_SEMITONE, NUM_SEMITONES,
with_bar_dependent_probs, hopTime, usul_type)
self.hmm.build_trans_probs(with_bar_dependent_probs)
self.hmm.build_obs_model()
def path_to_stepstates(self, path):
'''
convert path to a list of FrameOutput with 3 states : 1: attack, 2: sustain, 3: silence
'''
out = []
for iFrame in range(len(path)):
currPitch = -1.0
stateKind = 0
currPitch = self.hmm.par.minPitch + (
path[iFrame] / self.hmm.par.nSPP) * 1.0 / self.hmm.par.nPPS
stateKind = (
path[iFrame]
) % self.hmm.par.nSPP + 1 # 1: attack, 2: sustain, 3: silence
out.append(FrameOutput(iFrame, currPitch, stateKind))
return out
def process(self, pitch_contour_and_prob, bar_position_ts, bar_labels,
hop_time):
'''
compute obs. probabilities and decode with Viterbi
Parameters
----------------------
pitch_contour_and_prob
pitch observation feature
bar_position_ts: list
timestamps of bar positions
bar_labels: list
labels of bar types corresponding to timestamps
'''
obs_probs = self.hmm.calculatedObsProb(pitch_contour_and_prob)
obs_probs = self.hmm.normalize_obs_probs(obs_probs,
pitch_contour_and_prob)
obs_probs_T = obs_probs.T
self.create_beatPositions(obs_probs_T, bar_position_ts, bar_labels,
hop_time)
path, _ = self.hmm.decodeViterbi(
obs_probs_T) # transpose to have time t as first dimension
out = self.path_to_stepstates(path)
return out
def create_beatPositions(self, obs_probs_T, beat_position_ts, beat_labels,
hop_time):
'''
load beat annotaiton. and create beat position markers at frames
creates MonoNoteHMM.beatPositions: shape(time, 2);
dimension 0: one if no bar pos, else zero;
dimension 1: bar pos label (form 0 to num_beats in usul)
'''
nFrames = obs_probs_T.shape[0]
self.hmm.beatPositions = np.zeros((nFrames, 2)) # create output
for beat_pos_ts, beat_label in zip(beat_position_ts, beat_labels):
iFrame = ts_to_frame(beat_pos_ts, hop_time)
if iFrame >= nFrames - 1:
logging.warning(
'bar position ts beyond duration of audio... ignoring')
break
self.hmm.beatPositions[iFrame, 0] = 1
self.hmm.beatPositions[iFrame, 1] = int(beat_label) - 1
