def get_last_beat(cumscore):
"""Get the last beat from the cumulative score array"""
maxes = librosa.localmax(cumscore)
med_score = np.median(cumscore[np.argwhere(maxes)])
# The last of these is the last beat (since score generally increases)
return np.argwhere((cumscore * maxes * 2 > med_score)).max()
def onset_estimate_bpm(onsets, start_bpm, fft_res):
"""Estimate the BPM from an onset envelope
Arguments:
onsets -- (ndarray) time-series of onset strengths
start_bpm -- (float) initial guess of the BPM
fft_res -- (float) resolution of FFT (sample rate / hop length)
Returns bpm:
bpm -- (float) estimated BPM
"""
ac_size = 4.0
duration = 90.0
end_time = 90.0
bpm_std = 1.0
# Chop onsets to X[(upper_limit - duration):upper_limit]
# or as much as will fit
maxcol = min(len(onsets)-1, np.round(end_time * fft_res))
mincol = max(0, maxcol - np.round(duration * fft_res))
# Use auto-correlation out of 4 seconds (empirically set??)
ac_window = np.round(ac_size * fft_res)
# Compute the autocorrelation
x_corr = librosa.autocorrelate(onsets[mincol:maxcol], ac_window)
# FIXME: 2013-01-25 08:55:40 by Brian McFee <*****@*****.**>
# this fails if ac_window > length of song
# re-weight the autocorrelation by log-normal prior
bpms = 60.0 * fft_res / (np.arange(1, ac_window+1))
# Smooth the autocorrelation by a log-normal distribution
x_corr = x_corr * np.exp(-0.5 * ((np.log2(bpms / start_bpm)) / bpm_std)**2)
# Get the local maximum of weighted correlation
x_peaks = librosa.localmax(x_corr)
# Zero out all peaks before the first negative
x_peaks[:np.argmax(x_corr < 0)] = False
# Choose the best peak out of .33, .5, 2, 3 * start_period
candidates = np.multiply( np.argmax(x_peaks * x_corr),
[1.0/3, 1.0/2, 1.0, 2.0, 3.0])
candidates = candidates.astype(int)
candidates = candidates[candidates < ac_window]
best_period = np.argmax(x_corr[candidates])
return 60.0 * fft_res / candidates[best_period]
def activation_upsample(A_low, sr):
n, k, _, sr_old = A_low.shape
A = np.zeros((n, k, 1, sr))
for i in range(n):
for j in range(k):
act_res = librosa.resample(A_low[i, j, :, :].squeeze(), orig_sr=sr_old, target_sr=sr)
# Local-max filter act_res
A[i, j, :, : min(sr, len(act_res))] = librosa.localmax(act_res) * act_res
return A
def activation_upsample(A_low, sr):
n, k, _, sr_old = A_low.shape
A = np.zeros((n, k, 1, sr))
for i in range(n):
for j in range(k):
act_res = librosa.resample(A_low[i, j, :, :].squeeze(),
orig_sr=sr_old,
target_sr=sr)
# Local-max filter act_res
A[i, j, :, :min(sr, len(act_res)
)] = librosa.localmax(act_res) * act_res
return A
def __test_peaks(tempo, win_length, window, norm):
# Generate an evenly-spaced pulse train
odf = np.zeros(duration * sr // hop_length)
spacing = sr * 60. // (hop_length * tempo)
odf[::int(spacing)] = 1
tempogram = librosa.feature.tempogram(onset_envelope=odf,
sr=sr,
hop_length=hop_length,
win_length=win_length,
window=window,
norm=norm)
# Check the shape of the output
eq_(tempogram.shape[0], win_length)
eq_(tempogram.shape[1], len(odf))
# Mean over time to wash over the boundary padding effects
idx = np.where(librosa.localmax(tempogram.max(axis=1)))[0]
# Indices should all be non-zero integer multiples of spacing
assert np.allclose(idx, spacing * np.arange(1, 1 + len(idx)))
