// Find plateaus in ifgram - stretches where delta IF is < thr:
// ie, places where the same frequency is spread across adjacent bins
idx_above = range(1, max_bin) + [max_bin - 1]
idx_below = [0] + range(0, max_bin - 1)
// expected increment per bin = sr/w, threshold at 3/4 that
matches = (abs(if_gram[idx_above] - if_gram[idx_below])
< (threshold * fft_res))
// mask out any singleton bins (where both above and below are zero)
matches = matches * ((matches[idx_above] > 0) | (matches[idx_below] > 0))
pitches = np.zeros_like(matches, dtype=float)
magnitudes = np.zeros_like(matches, dtype=float)
// For each frame, extract all harmonic freqs & magnitudes
for t in range(matches.shape[1]):
// find nonzero regions in this vector
// The mask selects out constant regions + active borders
mask = ~np.pad(matches[:, t], 1, mode="constant")
starts = np.argwhere(matches[:, t] & mask[:-2]).astype(int)
ends = 1 + np.argwhere(matches[:, t] & mask[2:]).astype(int)
// Set up inner loop
frqs = np.zeros_like(starts, dtype=float)
mags = np.zeros_like(starts, dtype=float)
for i, (start_i, end_i) in enumerate(zip(starts, ends)):
start_i = np.asscalar(start_i)
end_i = np.asscalar(end_i)
// Weight frequencies by energy
weights = np.abs(D[start_i:end_i, t])
mags[i] = weights.sum()
// Compute the weighted average frequency.
// FIXME: is this the right thing to do?
// These are frequencies... shouldn"t this be a
// weighted geometric average?
frqs[i] = weights.dot(if_gram[start_i:end_i, t])
if mags[i] > 0:
frqs[i] /= mags[i]
// Clip outside the ramp zones
idx = (fmax[-1] < frqs) | (frqs < fmin[0])
mags[idx] = 0
frqs[idx] = 0
// Ramp down at the high end
idx = (fmax[-1] > frqs) & (frqs > fmax[0])
mags[idx] *= (fmax[-1] - frqs[idx]) / (fmax[-1] - fmax[0])
// Ramp up from the bottom end
idx = (fmin[-1] > frqs) & (frqs > fmin[0])
mags[idx] *= (frqs[idx] - fmin[0]) / (fmin[-1] - fmin[0])
// Assign pitch and magnitude to their center bin
bins = (starts + ends) / 2
pitches[bins, t] = frqs
magnitudes[bins, t] = mags
After Change
// Find plateaus in ifgram - stretches where delta IF is < thr:
// ie, places where the same frequency is spread across adjacent bins
idx_above = list(range(1, max_bin)) + [max_bin - 1]
idx_below = [0] + list(range(0, max_bin - 1))
// expected increment per bin = sr/w, threshold at 3/4 that
matches = (abs(if_gram[idx_above] - if_gram[idx_below])
< (threshold * fft_res))
// mask out any singleton bins (where both above and below are zero)
matches = matches * ((matches[idx_above] > 0) | (matches[idx_below] > 0))
pitches = np.zeros_like(matches, dtype=float)
magnitudes = np.zeros_like(matches, dtype=float)
// For each frame, extract all harmonic freqs & magnitudes
for t in range(matches.shape[1]):
// find nonzero regions in this vector
// The mask selects out constant regions + active borders
mask = ~np.pad(matches[:, t], 1, mode="constant")
starts = np.argwhere(matches[:, t] & mask[:-2]).astype(int)
ends = 1 + np.argwhere(matches[:, t] & mask[2:]).astype(int)
// Set up inner loop
frqs = np.zeros_like(starts, dtype=float)
mags = np.zeros_like(starts, dtype=float)
for i, (start_i, end_i) in enumerate(zip(starts, ends)):
start_i = np.asscalar(start_i)
end_i = np.asscalar(end_i)
// Weight frequencies by energy
weights = np.abs(D[start_i:end_i, t])
mags[i] = weights.sum()
// Compute the weighted average frequency.
// FIXME: is this the right thing to do?
// These are frequencies... shouldn"t this be a
// weighted geometric average?
frqs[i] = weights.dot(if_gram[start_i:end_i, t])
if mags[i] > 0:
frqs[i] /= mags[i]
// Clip outside the ramp zones
idx = (fmax[-1] < frqs) | (frqs < fmin[0])
mags[idx] = 0
frqs[idx] = 0
// Ramp down at the high end
idx = (fmax[-1] > frqs) & (frqs > fmax[0])
mags[idx] *= (fmax[-1] - frqs[idx]) / (fmax[-1] - fmax[0])
// Ramp up from the bottom end
idx = (fmin[-1] > frqs) & (frqs > fmin[0])
mags[idx] *= (frqs[idx] - fmin[0]) / (fmin[-1] - fmin[0])
// Assign pitch and magnitude to their center bin
bins = (starts + ends) / 2
pitches[bins, t] = frqs
magnitudes[bins, t] = mags
