The sentence
The sentence, for reference, with Harvest's track drawn on the spectrogram.
vc1_the_sentence.flac, 2.5 s, mono
Code
sound = sentence
fig, playhead = show(sound, "the sentence", harvest)Code
import matplotlib.pyplot as plt
import numpy as np
import sonore as so
plt.rcParams.update({"font.size": 9, "axes.titlesize": 10, "figure.dpi": 100})
def finish(snd):
"""How every sound in the gallery is played: 5 ms ramps, RMS 0.1, peak at most 0.95."""
snd = snd.ramp(5e-3).normalize(rms=0.1)
return snd.normalize(peak=0.95) if snd.peak > 0.95 else snd
def show(snd, title, contour=None, fmax=5000):
"""Waveform and narrowband spectrogram (Hann 33 ms, harmonics resolved), on one time axis,
with an F0 contour drawn over the spectrogram if one is given."""
fig = plt.figure(figsize=(10, 4.2), layout="constrained")
ax0, ax1 = fig.subplots(2, 1, sharex=True, height_ratios=[0.45, 1])
snd.plot(ax0, color="k", lw=0.4)
ax0.set(title=f"Waveform: {title}", xlabel="")
so.STFT(snd, win_dur=0.0333, hop_dur=0.002).plot(ax1, db_range=70, colorbar=False, fmax=fmax)
ax1.set_title("Spectrogram (Hann 33 ms)")
if contour is not None:
times, values = contour
ax1.plot(times, np.where(values > 0, values, np.nan), color="c", lw=1.2)
for ax in (ax0, ax1):
ax.set_xlim(0, snd.duration)
return fig, [ax0, ax1]
sentence = finish(so.load("docs/speech/bdl_arctic_a0131.flac"))
fs = sentence.fs
harvest_times, harvest_f0 = np.loadtxt("docs/speech/bdl_arctic_a0131_f0.csv", delimiter=",", skiprows=2).T
harvest = (harvest_times, harvest_f0)
envelope = so.cheaptrick(sentence, harvest)
aperiodicity = so.d4c(sentence, harvest)
voiced = harvest_f0 > 0
scored = (envelope.f >= 100) & (envelope.f <= 5000)
def median_f0(snd):
"""The median F0 that so.f0_track measures over the voiced time windows of a sound."""
track = so.f0_track(snd)
return np.median(track.f0[0][track.voiced[0]])
def mean_voiced_db(view, times, is_voiced):
"""An envelope averaged in dB over the voiced time windows, on CheapTrick's frequencies."""
return np.mean(10 * np.log10(view(times[is_voiced], envelope.f)[0]), axis=1)
def misfit(changed_db, reference_db, ratio, slope=False):
"""How far changed_db is from reference_db read at f / ratio (RMS dB, 100-5000 Hz), with
the level left free and, if slope is True, a straight line in dB over log frequency too."""
difference = changed_db[scored] - np.interp(envelope.f[scored] / ratio, envelope.f, reference_db)
columns = [np.ones(scored.sum())] + ([np.log2(envelope.f[scored])] if slope else [])
fit = np.linalg.lstsq(np.column_stack(columns), difference, rcond=None)[0]
return np.sqrt(np.mean((difference - np.column_stack(columns) @ fit) ** 2))
RATIOS = np.arange(0.70, 2.001, 0.005)
def fitted_warp(changed_db, reference_db, slope=False):
"""The ratio r for which changed_db best matches reference_db read at f / r."""
misfits = [misfit(changed_db, reference_db, ratio, slope) for ratio in RATIOS]
return RATIOS[np.argmin(misfits)], np.min(misfits)
original_f0 = median_f0(sentence)
original_db = mean_voiced_db(envelope, harvest_times, voiced)
def report(snd, track):
"""Print the measured pitch ratio and fitted warp of a changed sentence; track is its F0
contour."""
changed_db = mean_voiced_db(so.cheaptrick(snd, track), harvest_times, voiced)
warp = fitted_warp(changed_db, original_db)[0]
print(f"pitch ratio {median_f0(snd) / original_f0:.3f}, fitted warp {warp:.3f}")
print(f"median voiced F0 of the sentence: {original_f0:.0f} Hz")median voiced F0 of the sentence: 119 Hz
