// Search range for previous beat: number of samples forward/backward to look
search_window = numpy.round(numpy.arange(-2 * period, -period/2 + 1))
// TODO: 2012-11-07 15:25:30 by Brian McFee <brm2132@columbia.edu>
// make a better variable name for this
// Make a score window, which begins biased toward start_bpm and skewed
//txwt = numpy.exp(-0.5 * (tightness * (numpy.log(-search_window) - numpy.log(period)))**2)
txwt = - tightness * numpy.abs(numpy.log(-search_window) - numpy.log(period))**2
// Are we on the first beat?
first_beat = True
// Forward step
for i in xrange(len(localscore)):
time_range = i + search_window
// Are we reaching back before time 0?
z_pad = numpy.maximum(0, numpy.minimum(- time_range[0], len(search_window)))
// max
// 0
// min
// -time_range[0] (2 * period - i) // for big enough i, this wins
// len(search_window) (3 * period / 2)
// Search over all possible predecessors and apply transition weighting
// txwt + [zeros(1, zpad), cumscore(time_range[zpad:])]
score_candidates = txwt.copy()
score_candidates[z_pad:] = score_candidates[z_pad:] + cumscore[time_range[z_pad:]]
// Find the best predecessor beat
beat_location = numpy.argmax(score_candidates)
current_score = score_candidates[beat_location]
// Add the local score
//cumscore[i] = alpha * current_score + (1-alpha) * localscore[i]
cumscore[i] = current_score + localscore[i] - alpha
// Special case the first onset. Stop if the localscore is small
if first_beat and localscore[i] < 0.01 * max_localscore:
backlink[i] = -1
else:
backlink[i] = time_range[beat_location]
first_beat = False
pass
pass
// TODO: 2012-11-07 16:30:23 by Brian McFee <brm2132@columbia.edu>
// pick up here
////// Get the last beat
maxes = librosa.localmax(cumscore)
max_indices = numpy.nonzero(maxes)[0]
peak_scores = cumscore[max_indices]
// TODO: 2012-11-07 16:29:29 by Brian McFee <brm2132@columbia.edu>
// what is this sorcery?
median_score = numpy.median(peak_scores)
bestendposs = numpy.nonzero(cumscore * maxes > 0.5 * median_score)[0]
// The last of these is the last beat (since score generally increases)
bestendx = numpy.max(bestendposs)
b = [int(bestendx)]
backlink = backlink.astype(int)
while backlink[b[-1]] >= 0:
b = numpy.concatenate((b, [backlink[b[-1]]]), axis=0)
pass
return b[::-1]
def onset_estimate_bpm(onsets, sampling_rate=8000, hop_length=32, start_bpm=120):
auto_correlation_size = 4.0
