def relative_local_speechrate(reft,
targett,
s=0.03,
realigntimes=False,
debug=False):
""" calculate the relative local speech rate between targett and
reft by DTW aligning the tracks, fitting a smoothing spline
to the frame time difference function (smoothing factor 's')
and using this to calculate derivative contour.
"""
try:
assert (targett.times[1] - targett.times[0]) == (
reft.times[1] - reft.times[0]
), "constant timestep for reference and target need to be equal..."
except AssertionError:
print(
"WARNING: timesteps must be equal.... (this may be spurious if 'remove_pau' was used)"
)
path = dtw_align(reft, targett)[-1]
ltd = _local_timediff(path, reft.times, targett.times)
if realigntimes:
newreftimes = np.arange(1,
len(ltd) + 1) * (ltd.times[1] - ltd.times[0])
else:
newreftimes = ltd.times.copy()
spline = UnivariateSpline(newreftimes, ltd.values, k=5, s=s)
if debug:
import pylab as pl
pl.plot(ltd.times, ltd.values)
pl.plot(ltd.times, spline(ltd.times))
rlstrack = Track()
rlstrack.times = newreftimes
rlstrack.values = spline(rlstrack.times, 1).reshape((-1, 1))
return rlstrack
def relative_local_x(reft, targett, dtwpath, s=100, debug=False):
"""Determines the rate of change of target relative to ref, using
alignment provided by dtwpath.
"""
t = reft.times.copy()
delta = np.zeros((len(t)), dtype=np.float64)
for rti, tti in dtwpath:
delta[rti] = targett.values[tti] - reft.values[rti]
spline = UnivariateSpline(t, delta, k=5, s=s)
if debug:
import pylab as pl
pl.subplot(211)
pl.plot(t, reft.values, label="ref")
temp = np.zeros((len(t)), dtype=np.float64)
for rti, tti in dtwpath:
temp[rti] = targett.values[tti]
pl.plot(t, temp, label="tgt")
pl.legend()
pl.subplot(212)
pl.plot(t, delta)
pl.plot(t, spline(t))
rlxtrack = Track()
rlxtrack.times = t
rlxtrack.values = spline(rlxtrack.times, 1).reshape((-1, 1))
return rlxtrack
def get_f0(args):
fn, f0_path, f0min, f0max, tstep, semitones, outf0dir = args
basename = os.path.basename(fn).split(".")[0]
print("PROCESSING: " + basename)
t = Track()
t.name = basename
t.get_f0(fn, f0min, f0max, timestep=tstep, semitones=semitones)
ttslab.tofile(t, os.path.join(outf0dir, basename + "." + TRACK_EXT))
def _local_timediff(path, reftimes, targettimes):
""" determine the local time difference mapped onto reftimes...
"""
t = reftimes[:path[-1][0] + 1]
delta = np.zeros((len(t)), dtype=np.float64)
for rti, tti in path:
delta[rti] = targettimes[tti] - reftimes[rti]
ttrack = Track()
ttrack.times = t
ttrack.values = delta
return ttrack
def utt_mceps(utt, shift=0.005, remove_pau=False, resettimes=False):
temppath = mkdtemp()
#wavs
wfn1 = os.path.join(temppath, "1." + WAV_EXT)
utt["waveform"].write(wfn1)
#feats
ffn1 = os.path.join(temppath, "1." + FEAT_EXT)
cmds = SIG2FV % {"inputfile": wfn1,
"outputfile": ffn1,
"shift": shift}
#print(cmds)
os.system(cmds)
#tracks
t1 = Track()
t1.load_track(ffn1)
#cleanup
shutil.rmtree(temppath)
keep_intervals = []
if remove_pau:
u = deepcopy(utt)
fill_startendtimes(u)
for seg in u.gr("Segment"):
if seg["name"] != "pau":
keep_intervals.append((seg["start"], seg["end"]))
indices = t1.mask_indices(keep_intervals)
t1.times = t1.times[indices]
t1.values = t1.values[indices]
if resettimes:
t1.times = np.arange(1, len(t1.times) + 1, dtype=np.float) * shift
return t1
def linearpath_distances(track, track2, metric="euclidean", VI=None):
dist = cdist(track.values, track2.values, metric=str(metric), VI=VI)
framedists = []
try:
for i in range(len(track.times)):
framedists.append(dist[i][i])
except IndexError:
pass
t = Track()
t.values = np.array(framedists)
t.values = t.values.reshape(-1, 1)
t.times = np.array([track.times[i] for i in range(len(t.values))])
if track2.numframes != track.numframes:
print("linearpath_distances: WARNING: num frames difference is %s" % (track2.numframes - track.numframes))
return t
def dtw_distances(track, track2, metric="euclidean", VI=None):
cumdist, dist, path = track.dtw_align(track2, metric=str(metric), VI=VI)
framedists = []
frametimes = []
for pathcoord in path:
x, y = pathcoord
framedists.append(dist[x][y])
frametimes.append(track.times[x])
t = Track()
t.values = np.array(framedists)
t.values = t.values.reshape(-1, 1)
t.times = np.array(frametimes)
return t
def dtw_distances(track, track2, metric="euclidean", VI=None):
cumdist, dist, path = track.dtw_align(track2, metric=str(metric), VI=VI)
framedists = []
frametimes = []
for pathcoord in path:
x, y = pathcoord
framedists.append(dist[x][y])
frametimes.append(track.times[x])
t = Track()
t.values = np.array(framedists)
t.values = t.values.reshape(-1, 1)
t.times = np.array(frametimes)
return t
def qta_synth_utt(utt, synthfunc=synth):
times = np.array([])
values = np.array([])
for phr in utt.gr("Phrase"):
synthparms = []
for word in phr.get_daughters():
for syl in word.gir("SylStructure").get_daughters():
synthparms.append([
syl[STARTLAB], syl[ENDLAB], syl[QTAPREFIX + "_endheight"],
syl[QTAPREFIX + "_slope"], syl[QTAPREFIX + "_lambd"]
])
phrf0track = synthfunc(phr[QTAPREFIX + "_startpitch"], synthparms)
times = np.concatenate((times, phrf0track.times))
values = np.concatenate((values, phrf0track.values.flatten()))
f0track = Track()
f0track.times = times
f0track.values = values.reshape((-1, 1))
return f0track
def gradient(track, h=2):
"""Estimate of the gradient using a window length 'h'... must be even... number of points is h+1
"""
assert h % 2 == 0
n = h // 2
#timesteps must be constant
period = track.times[1] - track.times[0]
times = track.times[n:-n].copy()
values = np.zeros(len(times), dtype=track.values.dtype)
for i in range(len(values)):
values[i] = (track.values[i + h] - track.values[i]) / (h * period)
t = Track()
t.times = times
t.values = values
return t
def synth(self, voice, utt, args):
synthparms = args #not yet implemented...
htslabel = "\n".join(utt["hts_label"]).encode(
"utf-8").splitlines() #to utf-8 bytestring
if synthparms and "use_labalignments" in synthparms:
use_labalignments = True
else:
use_labalignments = False
with HTS_EngineME(self.htsvoice_bin, self.mixfilter_bin,
self.pdfilter_bin) as htsengine:
htsengine.synth(htslabel, use_labalignments=use_labalignments)
for segt, seg in zip(htsengine.get_segtimes(), utt.gr("Segment")):
seg["start"], seg["end"] = segt
# utt["debug_waveform"] = htsengine.get_wav()
f0st = 12.0 * np.log2(htsengine.get_f0())
f0st[f0st == -np.inf] = 0.0
f0times = np.arange(len(f0st)) * STEPSIZE
f0track = Track()
f0track.times = f0times
f0track.values = f0st.reshape((-1, 1))
# utt["debug_f0track"] = f0track
utt.fill_startendtimes()
#add qta_startpitch
for phr in utt.gr("Phrase"):
syl = phr.first_daughter.gir("SylStructure").first_daughter
syltrackvals = f0track.slice(f0track.index_at(syl["start"]),
f0track.index_at(
syl["end"])).values.flatten()
validvals = syltrackvals[syltrackvals.nonzero()]
if len(validvals) > 3:
phr["qta_startpitch"] = np.mean(
validvals[:len(validvals) // 4])
else:
phr["qta_startpitch"] = BACKOFF_STARTPITCH
utt = voice.pitchmodel(utt, ("synth", None))
f0spline = InterpolatedUnivariateSpline(utt["f0track"].times,
utt["f0track"].values)
newf0 = f0spline(f0track.times)
#HEURISTIC ADJUSTMENT CLOSER TO HTS DYNAMICS
m = np.mean(f0track.values[f0track.values.nonzero()])
newf0 *= 1.3 #more dynamic
m2 = np.mean(newf0)
newf0 += m - m2
### TRANSFER UNVOICED SECTIONS
# newf0[f0track.values.flatten() == 0.0] = 0.0
# import pylab as pl
# pl.plot(f0track.times, f0track.values)
# pl.plot(f0track.times, newf0)
###
newf0 = 2**(newf0 / 12.0)
newf0 = tolf0(newf0)
htsengine.synth(htslabel,
lf0=newf0,
use_labalignments=use_labalignments)
#populate utt with waveform and segment alignments
utt["waveform"] = htsengine.get_wav()
return utt
def synth(startpitch, synthparms, numpoints=100, plot=False):
times = np.zeros(len(synthparms) * numpoints)
contour = np.zeros(len(synthparms) * numpoints)
for i, synthparm in enumerate(synthparms):
if i == 0:
p0 = startpitch
dp0 = 0.0
ddp0 = 0.0
if synthparm[0] != synthparms[
i - 1][1]: #not contiguous (e.g. a pause is present)
dp0 = 0.0
ddp0 = 0.0
if any([e is None for e in synthparm
]): #no parameters available for this syllable, skip...
dp0 = 0.0
ddp0 = 0.0
continue
utt_t = np.linspace(synthparm[0],
synthparm[1],
numpoints,
endpoint=False)
times[i * numpoints:i * numpoints + numpoints] = utt_t
syl_t = utt_t - synthparm[0] #start at 0.0
#y = mx + c
syltarget_m = synthparm[3]
syltarget_c = get_intercept(syl_t[-1], synthparm[2], synthparm[3])
scontour = sylcontour(syl_t, syltarget_m, syltarget_c, p0, dp0, ddp0,
synthparm[4])
if plot:
pl.plot(syl_t + synthparm[0],
np.polyval(coefs, syl_t),
linestyle="dashed",
color="red")
pl.plot(syl_t + synthparm[0], scontour, color="green")
spline = InterpolatedUnivariateSpline(syl_t, scontour)
contour[i * numpoints:i * numpoints + numpoints] = scontour
p0, dp0, ddp0, temp = spline.derivatives(syl_t[-1])
synthtrack = Track()
synthtrack.times = times[contour.nonzero()].copy()
synthtrack.values = contour[contour.nonzero()].reshape((-1, 1)).copy()
return synthtrack
def make_joincoefs(featconfig, wav_dir):
""" Make joincoefs...
"""
mcep_dir = os.path.join(os.getcwd(), MCEP_DIR)
os.mkdir(mcep_dir)
join_dir = os.path.join(os.getcwd(), JOIN_DIR)
os.mkdir(join_dir)
pm_dir = os.path.join(os.getcwd(), PM_DIR)
f0_dir = os.path.join(os.getcwd(), F0_DIR)
fbank_order = featconfig.get("SIG2FV_MCEP", "FBANK_ORDER")
melcep_order = featconfig.get("SIG2FV_MCEP", "MELCEP_ORDER")
melcep_coefs = featconfig.get("SIG2FV_MCEP", "MELCEP_COEFS")
preemph_coef = featconfig.get("SIG2FV_MCEP", "PREEMPH_COEF")
window_factor = featconfig.get("SIG2FV_MCEP", "WINDOW_FACTOR")
window_type = featconfig.get("SIG2FV_MCEP", "WINDOW_TYPE")
print("MAKING JOINCOEFS...")
map(extract_mceps,
[(wavfilename, fbank_order, window_factor, preemph_coef, melcep_order, window_type, melcep_coefs, mcep_dir, pm_dir)
for wavfilename in sorted(glob(os.path.join(wav_dir, ".".join(["*", WAV_EXT]))))])
print("NORMALISING AND JOINING F0 AND MCEPS...")
#Normalising mceps and f0s:
upper = +1.0
lower = -1.0
mceptracks = {}
for fn in glob(os.path.join(mcep_dir, ".".join(["*", MCEP_EXT]))):
t = Track()
t.load_track(fn)
mceptracks[os.path.basename(fn)] = t
allmcepvecs = np.concatenate([mceptracks[tn].values for tn in sorted(mceptracks)])
mcepmean = allmcepvecs.mean(0)
mcepstd = allmcepvecs.std(0)
for k in mceptracks:
mceptracks[k].values = (mceptracks[k].values - mcepmean) / (4 * mcepstd) * (upper - lower)
f0tracks = {}
for fn in glob(os.path.join(f0_dir, ".".join(["*", F0_EXT]))):
t = Track()
t.load_track(fn)
f0tracks[os.path.basename(fn)] = t
#allf0vecs = np.concatenate([f0tracks[tn].values for tn in sorted(f0tracks)])
allf0vecs = np.concatenate([f0tracks[tn].values[f0tracks[tn].values.nonzero()] for tn in sorted(f0tracks)])
f0mean = allf0vecs.mean(0)
f0std = allf0vecs.std(0)
for k in f0tracks:
f0tracks[k].values = (f0tracks[k].values - f0mean) / (4 * f0std) * (upper - lower)
#Add f0 to mcep track:
for k1, k2 in zip(sorted(mceptracks), sorted(f0tracks)):
mceptracks[k1].values = np.concatenate((mceptracks[k1].values, f0tracks[k2].values), 1)
for fn in mceptracks:
basename = os.path.splitext(os.path.basename(fn))[0]
ttslab.tofile(mceptracks[fn], os.path.join(join_dir, basename + "." + JOIN_EXT))
def timenorm_tonecontour_dtw(reftrack, track):
#normalise location:
tmean = deepcopy(reftrack)
tmean.values = tmean.values - tmean.values.mean()
#normalise location:
t1 = deepcopy(track)
t1.values = t1.values - t1.values.mean()
#smooth function to facilitate smoother warping:
s = UnivariateSpline(t1.times, t1.values)
t1.values = s(t1.times).reshape((-1, 1))
#align:
dtw = dtw_align(tmean, t1)
#construct new track using mapping:
newtrack = Track()
newtrack.times = np.copy(tmean.times)
values = np.zeros(len(tmean), np.float64)
for i in range(len(tmean)):
es = [e[1] for e in dtw[2] if e[0] == i]
values[i] = np.mean(track.values[es])
newtrack.values = values.reshape((-1, 1))
smoothtrack = newtrack.newtrack_from_sspline(newtrack.times,
s=len(newtrack.times) / 10.0)
return smoothtrack
def utt_distance(utt,
utt2,
method="dtw",
metric="euclidean",
sig2fv=SIG2FV,
VI=None):
""" Uses Trackfile class' distance measurements to compare utts...
See docstring in tfuncs_analysis.py for more details...
"""
temppath = mkdtemp()
#wavs
wfn1 = os.path.join(temppath, "1." + WAV_EXT)
wfn2 = os.path.join(temppath, "2." + WAV_EXT)
utt["waveform"].write(wfn1)
utt2["waveform"].write(wfn2)
#feats
ffn1 = os.path.join(temppath, "1." + FEAT_EXT)
ffn2 = os.path.join(temppath, "2." + FEAT_EXT)
cmds = SIG2FV % {"inputfile": wfn1, "outputfile": ffn1}
#print(cmds)
os.system(cmds)
cmds = SIG2FV % {"inputfile": wfn2, "outputfile": ffn2}
#print(cmds)
os.system(cmds)
#tracks
t1 = Track()
t1.load_track(ffn1)
t2 = Track()
t2.load_track(ffn2)
#compare and save
t3 = t1.distances(t2, method=method, metric=metric, VI=VI)
shutil.rmtree(temppath)
return t3
def utt_distance(utt, utt2, method="dtw", metric="euclidean", sig2fv=SIG2FV, VI=None):
""" Uses Trackfile class' distance measurements to compare utts...
See docstring in tfuncs_analysis.py for more details...
"""
temppath = mkdtemp()
#wavs
wfn1 = os.path.join(temppath, "1." + WAV_EXT)
wfn2 = os.path.join(temppath, "2." + WAV_EXT)
utt["waveform"].write(wfn1)
utt2["waveform"].write(wfn2)
#feats
ffn1 = os.path.join(temppath, "1." + FEAT_EXT)
ffn2 = os.path.join(temppath, "2." + FEAT_EXT)
cmds = SIG2FV % {"inputfile": wfn1,
"outputfile": ffn1}
#print(cmds)
os.system(cmds)
cmds = SIG2FV % {"inputfile": wfn2,
"outputfile": ffn2}
#print(cmds)
os.system(cmds)
#tracks
t1 = Track()
t1.load_track(ffn1)
t2 = Track()
t2.load_track(ffn2)
#compare and save
t3 = t1.distances(t2, method=method, metric=metric, VI=VI)
shutil.rmtree(temppath)
return t3
def add_feats_to_utt(args):
u, lpc_dir, joincoef_dir, f0_dir = args
file_id = u["file_id"]
print("Processing:", file_id)
u.fill_startendtimes()
for unit, word in zip(u.gr("Unit"), u.gr("Word")):
assert unit["name"] == word["name"]
unit["start"] = word["start"]
unit["end"] = word["end"]
lpctrack = Track()
lpctrack.load_track(".".join([os.path.join(lpc_dir, file_id), LPC_EXT]))
restrack = Track()
restrack.load_wave(".".join([os.path.join(lpc_dir, file_id), RES_EXT]))
jointrack = ttslab.fromfile(".".join([os.path.join(joincoef_dir, file_id), JOIN_EXT]))
f0track = Track()
f0track.load_track(".".join([os.path.join(f0_dir, file_id), F0_EXT]))
#get boundarytimes:
boundarytimes = []
for i, unit in enumerate(u.gr("Unit")):
if i == 0:
boundarytimes.append(unit["start"])
boundarytimes.append(unit["end"])
#convert boundtimes into sample ranges:
lpcsampleranges = []
f0sampleranges = []
joinsamples = []
for bound in boundarytimes:
lpcsampleranges.append(lpctrack.index_at(bound))
f0sampleranges.append(f0track.index_at(bound))
joinsamples.append(jointrack.values[jointrack.index_at(bound)])
#get pitchperiods at lpc indices
lpctimes = np.concatenate(([0.0], lpctrack.times))
pitchperiod = np.diff(lpctimes)
units = u.get_relation("Unit").as_list()
assert len(units) == len(lpcsampleranges) - 1
for jc0, jc1, lti0, lti1, fti0, fti1, i in zip(joinsamples[:-1], joinsamples[1:],
lpcsampleranges[:-1], lpcsampleranges[1:],
f0sampleranges[:-1], f0sampleranges[1:],
units):
# print(i["name"], "lpctrack[%s:%s]" % (lti0, lti1), "len(lpctrack)=%s" % len(lpctrack))
i["left-joincoef"] = jc0
i["right-joincoef"] = jc1
i["lpc-coefs"] = lpctrack.slice(lti0, lti1, copy=True) #like python indexing/slicing
if lti0 == 0:
i["lpc-coefs"].starttime = 0.0
else:
i["lpc-coefs"].starttime = lpctrack.times[lti0 - 1]
i["lpc-coefs"].zero_starttime()
#For windowfactor=2 (save only samples and assume 16kHz)
i["residuals"] = restrack.slice(restrack.index_at(lpctrack.times[lti0] - pitchperiod[lti0]),
restrack.index_at(lpctrack.times[lti1] + pitchperiod[lti0])).values
return u
def synth2(startpitch,
synthparms,
numpoints=100,
plot=False,
minlambd=10.0,
dlambd=5.0):
""" Limit the strength of articulation to avoid acceleration in
opposite direction of endheight target...
"""
times = np.zeros(len(synthparms) * numpoints)
contour = np.zeros(len(synthparms) * numpoints)
for i, synthparm in enumerate(synthparms):
if i == 0:
p0 = startpitch
dp0 = 0.0
ddp0 = 0.0
if synthparm[0] != synthparms[
i - 1][1]: #not contiguous (e.g. a pause is present)
dp0 = 0.0
ddp0 = 0.0
if any([e is None for e in synthparm
]): #no parameters available for this syllable, skip...
dp0 = 0.0
ddp0 = 0.0
continue
utt_t = np.linspace(synthparm[0],
synthparm[1],
numpoints,
endpoint=False)
times[i * numpoints:i * numpoints + numpoints] = utt_t
syl_t = utt_t - synthparm[0] #start at 0.0
#y = mx + c
syltarget_m = synthparm[3]
syltarget_c = get_intercept(syl_t[-1], synthparm[2], synthparm[3])
while True: #resynthesise with lower strength until constraint met
scontour = sylcontour(syl_t, syltarget_m, syltarget_c, p0, dp0,
ddp0, synthparm[4])
spline = InterpolatedUnivariateSpline(syl_t, scontour)
#check acceleration
if synthparm[4] <= minlambd:
break
accels = spline(syl_t, 2)
if synthparm[2] > p0:
if np.all(accels > 0.0):
break
elif synthparm[2] < p0:
if np.all(accels < 0.0):
break
else:
break
synthparm[4] -= dlambd
if synthparm[4] < minlambd:
synthparm[4] = minlambd
if plot:
pl.plot(syl_t + synthparm[0],
np.polyval(coefs, syl_t),
linestyle="dashed",
color="red")
pl.plot(syl_t + synthparm[0], scontour, color="green")
contour[i * numpoints:i * numpoints + numpoints] = scontour
p0, dp0, ddp0, temp = spline.derivatives(syl_t[-1])
synthtrack = Track()
synthtrack.times = times[contour.nonzero()].copy()
synthtrack.values = contour[contour.nonzero()].reshape((-1, 1)).copy()
return synthtrack
def draw_sylstruct_graph_pitch_waveform(u):
#use seg end times to calculate start and end times for all
#items...
u.fill_startendtimes()
g = nx.Graph()
posdict = {}
nodelist = []
nodesizelist = []
for word in u.get_relation("SylStructure"):
nodelist.append(word)
nodesizelist.append(300 * len(str(word)))
posdict[word] = [word["end"] + word["start"] / 2, 3]
if word.prev_item:
g.add_edge(word.prev_item, word)
if word.next_item:
g.add_edge(word.next_item, word)
g.add_edge(word.first_daughter, word)
g.add_edge(word.last_daughter, word)
for syl in word.get_daughters():
nodelist.append(syl)
nodesizelist.append(400)
posdict[syl] = [syl["end"] + syl["start"] / 2, 2]
if syl.prev_item:
g.add_edge(syl.prev_item, syl)
if syl.next_item:
g.add_edge(syl.next_item, syl)
g.add_edge(syl.first_daughter, syl)
g.add_edge(syl.last_daughter, syl)
for seg in syl.get_daughters():
nodelist.append(seg)
nodesizelist.append(350)
posdict[seg] = [seg["end"] + seg["start"] / 2, 1]
if seg.prev_item:
g.add_edge(seg.prev_item, seg)
if seg.next_item:
g.add_edge(seg.next_item, seg)
uttendtime = u.get_relation("Segment").tail_item["end"]
bounds = np.array([word["end"] for word in u.get_relation("Word")])
#get the pitch:
d = mkdtemp()
u["waveform"].write(os.path.join(d, "utt.wav"))
f0t = Track()
f0t.get_f0(os.path.join(d, "utt.wav"), semitones=True)
shutil.rmtree(d)
fig1 = plt.figure(
) #edgecolor=(0.921568627451, 0.921568627451, 0.921568627451))
ax = fig1.add_subplot(111)
ax.set_title("Utterance")
# ax.set_ylim(0, 5)
nx.draw(g, pos=posdict, ax=ax, nodelist=nodelist, node_size=nodesizelist)
plt.xticks([], [])
plt.yticks([1.0, 2.0, 3.0], ["segment", "syllable", "word"])
fig2 = plt.figure()
ax1 = fig2.add_subplot(111)
ax1.set_title("Pitch")
ax1.set_ylabel("Semitones (relative to 1 Hz)")
ax1.set_xlabel("Syllables")
plt.plot(f0t.times, f0t.values, color='green')
ax1.set_ylim(bottom=75.0)
plt.xticks([syl["end"] for syl in u.gr("Syllable")],
[getsylsegstr(syl) for syl in u.gr("Syllable")])
ax1.grid()
fig3 = plt.figure()
ax2 = fig3.add_subplot(111)
decimate_factor = 10
ax2.set_title("Waveform (decimation factor: %s)" % decimate_factor)
ax2.set_ylabel("Amplitude")
ax2.set_xlabel("Syllables")
waveform = ss.decimate(u["waveform"].samples, decimate_factor)
plt.plot(np.arange(len(waveform)) *
(1.0 / u["waveform"].samplerate * decimate_factor),
waveform,
color='b')
#ax2.set_xticks(bounds*u["waveform"].samplerate, [''] * len(bounds))
plt.xticks([syl["end"] for syl in u.gr("Syllable")],
[getsylsegstr(syl) for syl in u.gr("Syllable")])
# fig3.set_facecolor((0.921568627451, 0.921568627451, 0.921568627451))
ax2.grid()
#plt.show()
return fig1, fig2, fig3
def draw_sylstruct_graph_pitch_waveform(u):
#use seg end times to calculate start and end times for all
#items...
u.fill_startendtimes()
g = nx.Graph()
posdict = {}
nodelist = []
nodesizelist = []
for word in u.get_relation("SylStructure"):
nodelist.append(word)
nodesizelist.append(300 * len(str(word)))
posdict[word] = [word["end"] + word["start"] / 2, 3]
if word.prev_item:
g.add_edge(word.prev_item, word)
if word.next_item:
g.add_edge(word.next_item, word)
g.add_edge(word.first_daughter, word)
g.add_edge(word.last_daughter, word)
for syl in word.get_daughters():
nodelist.append(syl)
nodesizelist.append(400)
posdict[syl] = [syl["end"] + syl["start"] / 2, 2]
if syl.prev_item:
g.add_edge(syl.prev_item, syl)
if syl.next_item:
g.add_edge(syl.next_item, syl)
g.add_edge(syl.first_daughter, syl)
g.add_edge(syl.last_daughter, syl)
for seg in syl.get_daughters():
nodelist.append(seg)
nodesizelist.append(350)
posdict[seg] = [seg["end"] + seg["start"] / 2, 1]
if seg.prev_item:
g.add_edge(seg.prev_item, seg)
if seg.next_item:
g.add_edge(seg.next_item, seg)
uttendtime = u.get_relation("Segment").tail_item["end"]
bounds = np.array([word["end"] for word in u.get_relation("Word")])
#get the pitch:
d = mkdtemp()
u["waveform"].write(os.path.join(d, "utt.wav"))
f0t = Track()
f0t.get_f0(os.path.join(d, "utt.wav"))
shutil.rmtree(d)
fig = pl.figure(
) #edgecolor=(0.921568627451, 0.921568627451, 0.921568627451))
ax = fig.add_subplot(311)
ax.set_title("Utterance")
# ax.set_ylim(0, 5)
# ax.set_xticks(bounds)
# ax.grid()
nx.draw(g, pos=posdict, ax=ax, nodelist=nodelist, node_size=nodesizelist)
ax1 = fig.add_subplot(312)
ax1.set_title("Pitch")
ax1.set_ylim(20.0, 300.0)
ax1.set_ylabel("Hertz")
pl.plot(f0t.times, f0t.values, color='green')
pl.xticks([syl["end"] for syl in u.gr("Syllable")],
[syl["tone"] for syl in u.gr("Syllable")])
ax1.grid()
ax2 = fig.add_subplot(313)
ax2.set_title("Waveform")
ax2.set_xlim(0, uttendtime * u["waveform"].samplerate)
pl.plot(u["waveform"].samples, color='b')
ax2.set_xticks(bounds * u["waveform"].samplerate, [''] * len(bounds))
fig.set_facecolor((0.921568627451, 0.921568627451, 0.921568627451))
# ax2.grid()
# pl.show()
# fig.savefig("output.png")
return fig
def add_feats_to_utt(args):
u, lpc_dir, joincoef_dir, f0_dir = args
file_id = u["file_id"]
print("Processing:", file_id)
u.fill_startendtimes()
for unit, word in zip(u.gr("Unit"), u.gr("Word")):
assert unit["name"] == word["name"]
unit["start"] = word["start"]
unit["end"] = word["end"]
lpctrack = Track()
lpctrack.load_track(".".join([os.path.join(lpc_dir, file_id), LPC_EXT]))
restrack = Track()
restrack.load_wave(".".join([os.path.join(lpc_dir, file_id), RES_EXT]))
jointrack = ttslab.fromfile(".".join(
[os.path.join(joincoef_dir, file_id), JOIN_EXT]))
f0track = Track()
f0track.load_track(".".join([os.path.join(f0_dir, file_id), F0_EXT]))
#get boundarytimes:
boundarytimes = []
for i, unit in enumerate(u.gr("Unit")):
if i == 0:
boundarytimes.append(unit["start"])
boundarytimes.append(unit["end"])
#convert boundtimes into sample ranges:
lpcsampleranges = []
f0sampleranges = []
joinsamples = []
for bound in boundarytimes:
lpcsampleranges.append(lpctrack.index_at(bound))
f0sampleranges.append(f0track.index_at(bound))
joinsamples.append(jointrack.values[jointrack.index_at(bound)])
#get pitchperiods at lpc indices
lpctimes = np.concatenate(([0.0], lpctrack.times))
pitchperiod = np.diff(lpctimes)
units = u.get_relation("Unit").as_list()
assert len(units) == len(lpcsampleranges) - 1
for jc0, jc1, lti0, lti1, fti0, fti1, i in zip(joinsamples[:-1],
joinsamples[1:],
lpcsampleranges[:-1],
lpcsampleranges[1:],
f0sampleranges[:-1],
f0sampleranges[1:], units):
# print(i["name"], "lpctrack[%s:%s]" % (lti0, lti1), "len(lpctrack)=%s" % len(lpctrack))
i["left-joincoef"] = jc0
i["right-joincoef"] = jc1
i["lpc-coefs"] = lpctrack.slice(
lti0, lti1, copy=True) #like python indexing/slicing
if lti0 == 0:
i["lpc-coefs"].starttime = 0.0
else:
i["lpc-coefs"].starttime = lpctrack.times[lti0 - 1]
i["lpc-coefs"].zero_starttime()
#For windowfactor=2 (save only samples and assume 16kHz)
i["residuals"] = restrack.slice(
restrack.index_at(lpctrack.times[lti0] - pitchperiod[lti0]),
restrack.index_at(lpctrack.times[lti1] + pitchperiod[lti0])).values
return u
import sys
import array
import math
import numpy as np
import ttslab
from ttslab.trackfile import Track
ttslab.extend(Track, "ttslab.trackfile.funcs.tfuncs_praat")
def friendly_log(f):
try:
return math.log(f)
except ValueError:
return float('-1e+10')
if __name__ == "__main__":
fn = sys.argv[1]
outfn = sys.argv[2]
minf0 = float(sys.argv[3])
maxf0 = float(sys.argv[4])
t = Track()
t.get_f0(fn, minpitch=minf0, maxpitch=maxf0, timestep=0.005, fixocterrs=True) #timestep hardcoded here because of hack below...
#hack aligns samples with equiv from HTS script:
pad = np.array([0.0, 0.0]).reshape(-1, 1)
f0hzvalues = np.concatenate([pad, t.values, pad])
lf0 = array.array(b"f", map(friendly_log, f0hzvalues))
with open(outfn, "wb") as outfh:
lf0.tofile(outfh)
def add_feats_to_utt(args):
u, lpc_dir, joincoef_dir, f0_dir = args
file_id = u["file_id"]
print("Processing:", file_id)
u.fill_startendtimes()
lpctrack = Track()
lpctrack.load_track(".".join([os.path.join(lpc_dir, file_id), LPC_EXT]))
restrack = Track()
restrack.load_wave(".".join([os.path.join(lpc_dir, file_id), RES_EXT]))
jointrack = ttslab.fromfile(".".join([os.path.join(joincoef_dir, file_id), JOIN_EXT]))
f0track = Track()
f0track.load_track(".".join([os.path.join(f0_dir, file_id), F0_EXT]))
#get boundarytimes:
boundarytimes = []
durations = []
starttime = 0.0
for seg in u.get_relation("Segment"):
endtime = float(seg["end"])
if "cl_end" in seg:
splittime = float(seg["cl_end"])
else:
splittime = (endtime + starttime) / 2
#TODO: should still add 25% split if diphthong...
boundarytimes.append([starttime, splittime, endtime])
durations.extend([splittime - starttime, endtime - splittime])
starttime = endtime
#convert boundtimes into sample ranges (and flatten):
lpcsampleranges = []
f0sampleranges = []
joinsamples = []
#DEMITASSE: If not pruning pau halfphones:
# for bounds in boundarytimes:
# lpcsampleranges.extend([lpctrack.get_index_at(bounds[0]),
# lpctrack.get_index_at(bounds[1])])
# joinsamples.extend([jointrack.get_sample_at(bounds[0]),
# jointrack.get_sample_at(bounds[1])])
# lpcsampleranges.append(len(lpctrack))
# joinsamples.append(jointrack.get_sample_at(len(jointrack)))
#DEMITASSE: If pruning pau halfphones:
durations = durations[1:-1]
for i, bounds in enumerate(boundarytimes):
if i == 0:
lpcsampleranges.append(lpctrack.index_at(bounds[1]))
f0sampleranges.append(f0track.index_at(bounds[1]))
joinsamples.append(jointrack.values[bounds[1]])
else:
lpcsampleranges.extend([lpctrack.index_at(bounds[0]),
lpctrack.index_at(bounds[1])])
f0sampleranges.extend([f0track.index_at(bounds[0]),
f0track.index_at(bounds[1])])
joinsamples.extend([jointrack.values[bounds[0]],
jointrack.values[bounds[1]]])
#get pitchperiods at lpc indices
lpctimes = np.concatenate(([0.0], lpctrack.times))
pitchperiod = np.diff(lpctimes)
units = u.get_relation("Unit").as_list()
assert len(units) == len(lpcsampleranges) - 1
for jc0, jc1, lti0, lti1, fti0, fti1, dur, i in zip(joinsamples[:-1], joinsamples[1:],
lpcsampleranges[:-1], lpcsampleranges[1:],
f0sampleranges[:-1], f0sampleranges[1:],
durations,
units):
# print(i["name"], "lpctrack[%s:%s]" % (lti0, lti1), "len(lpctrack)=%s" % len(lpctrack))
i["left-joincoef"] = jc0
i["right-joincoef"] = jc1
i["lpc-coefs"] = lpctrack.slice(lti0, lti1, copy=True) #like python indexing/slicing
if lti0 == 0:
i["lpc-coefs"].starttime = 0.0
else:
i["lpc-coefs"].starttime = lpctrack.times[lti0 - 1]
i["lpc-coefs"].zero_starttime()
i["dur"] = dur
#For windowfactor=2 (save only samples and assume 16kHz)
i["residuals"] = restrack.slice(restrack.index_at(lpctrack.times[lti0] - pitchperiod[lti0]),
restrack.index_at(lpctrack.times[lti1] + pitchperiod[lti0])).values
return u
def make_joincoefs(featconfig, wav_dir):
""" Make joincoefs...
"""
mcep_dir = os.path.join(os.getcwd(), MCEP_DIR)
os.mkdir(mcep_dir)
join_dir = os.path.join(os.getcwd(), JOIN_DIR)
os.mkdir(join_dir)
pm_dir = os.path.join(os.getcwd(), PM_DIR)
f0_dir = os.path.join(os.getcwd(), F0_DIR)
fbank_order = featconfig.get("SIG2FV_MCEP", "FBANK_ORDER")
melcep_order = featconfig.get("SIG2FV_MCEP", "MELCEP_ORDER")
melcep_coefs = featconfig.get("SIG2FV_MCEP", "MELCEP_COEFS")
preemph_coef = featconfig.get("SIG2FV_MCEP", "PREEMPH_COEF")
window_factor = featconfig.get("SIG2FV_MCEP", "WINDOW_FACTOR")
window_type = featconfig.get("SIG2FV_MCEP", "WINDOW_TYPE")
print("MAKING JOINCOEFS...")
map(extract_mceps,
[(wavfilename, fbank_order, window_factor, preemph_coef, melcep_order,
window_type, melcep_coefs, mcep_dir, pm_dir)
for wavfilename in sorted(
glob(os.path.join(wav_dir, ".".join(["*", WAV_EXT]))))])
print("NORMALISING AND JOINING F0 AND MCEPS...")
#Normalising mceps and f0s:
upper = +1.0
lower = -1.0
mceptracks = {}
for fn in glob(os.path.join(mcep_dir, ".".join(["*", MCEP_EXT]))):
t = Track()
t.load_track(fn)
mceptracks[os.path.basename(fn)] = t
allmcepvecs = np.concatenate(
[mceptracks[tn].values for tn in sorted(mceptracks)])
mcepmean = allmcepvecs.mean(0)
mcepstd = allmcepvecs.std(0)
for k in mceptracks:
mceptracks[k].values = (mceptracks[k].values -
mcepmean) / (4 * mcepstd) * (upper - lower)
f0tracks = {}
for fn in glob(os.path.join(f0_dir, ".".join(["*", F0_EXT]))):
t = Track()
t.load_track(fn)
f0tracks[os.path.basename(fn)] = t
#allf0vecs = np.concatenate([f0tracks[tn].values for tn in sorted(f0tracks)])
allf0vecs = np.concatenate([
f0tracks[tn].values[f0tracks[tn].values.nonzero()]
for tn in sorted(f0tracks)
])
f0mean = allf0vecs.mean(0)
f0std = allf0vecs.std(0)
for k in f0tracks:
f0tracks[k].values = (f0tracks[k].values -
f0mean) / (4 * f0std) * (upper - lower)
#Add f0 to mcep track:
for k1, k2 in zip(sorted(mceptracks), sorted(f0tracks)):
mceptracks[k1].values = np.concatenate(
(mceptracks[k1].values, f0tracks[k2].values), 1)
for fn in mceptracks:
basename = os.path.splitext(os.path.basename(fn))[0]
ttslab.tofile(mceptracks[fn],
os.path.join(join_dir, basename + "." + JOIN_EXT))
def add_feats_to_utt(args):
u, lpc_dir, joincoef_dir, f0_dir = args
file_id = u["file_id"]
print("Processing:", file_id)
u.fill_startendtimes()
lpctrack = Track()
lpctrack.load_track(".".join([os.path.join(lpc_dir, file_id), LPC_EXT]))
restrack = Track()
restrack.load_wave(".".join([os.path.join(lpc_dir, file_id), RES_EXT]))
jointrack = ttslab.fromfile(".".join(
[os.path.join(joincoef_dir, file_id), JOIN_EXT]))
f0track = Track()
f0track.load_track(".".join([os.path.join(f0_dir, file_id), F0_EXT]))
#get boundarytimes:
boundarytimes = []
durations = []
starttime = 0.0
for seg in u.get_relation("Segment"):
endtime = float(seg["end"])
if "cl_end" in seg:
splittime = float(seg["cl_end"])
else:
splittime = (endtime + starttime) / 2
#TODO: should still add 25% split if diphthong...
boundarytimes.append([starttime, splittime, endtime])
durations.extend([splittime - starttime, endtime - splittime])
starttime = endtime
#convert boundtimes into sample ranges (and flatten):
lpcsampleranges = []
f0sampleranges = []
joinsamples = []
#DEMITASSE: If not pruning pau halfphones:
# for bounds in boundarytimes:
# lpcsampleranges.extend([lpctrack.get_index_at(bounds[0]),
# lpctrack.get_index_at(bounds[1])])
# joinsamples.extend([jointrack.get_sample_at(bounds[0]),
# jointrack.get_sample_at(bounds[1])])
# lpcsampleranges.append(len(lpctrack))
# joinsamples.append(jointrack.get_sample_at(len(jointrack)))
#DEMITASSE: If pruning pau halfphones:
durations = durations[1:-1]
for i, bounds in enumerate(boundarytimes):
if i == 0:
lpcsampleranges.append(lpctrack.index_at(bounds[1]))
f0sampleranges.append(f0track.index_at(bounds[1]))
joinsamples.append(jointrack.values[bounds[1]])
else:
lpcsampleranges.extend(
[lpctrack.index_at(bounds[0]),
lpctrack.index_at(bounds[1])])
f0sampleranges.extend(
[f0track.index_at(bounds[0]),
f0track.index_at(bounds[1])])
joinsamples.extend(
[jointrack.values[bounds[0]], jointrack.values[bounds[1]]])
#get pitchperiods at lpc indices
lpctimes = np.concatenate(([0.0], lpctrack.times))
pitchperiod = np.diff(lpctimes)
units = u.get_relation("Unit").as_list()
assert len(units) == len(lpcsampleranges) - 1
for jc0, jc1, lti0, lti1, fti0, fti1, dur, i in zip(
joinsamples[:-1], joinsamples[1:], lpcsampleranges[:-1],
lpcsampleranges[1:], f0sampleranges[:-1], f0sampleranges[1:],
durations, units):
# print(i["name"], "lpctrack[%s:%s]" % (lti0, lti1), "len(lpctrack)=%s" % len(lpctrack))
i["left-joincoef"] = jc0
i["right-joincoef"] = jc1
i["lpc-coefs"] = lpctrack.slice(
lti0, lti1, copy=True) #like python indexing/slicing
if lti0 == 0:
i["lpc-coefs"].starttime = 0.0
else:
i["lpc-coefs"].starttime = lpctrack.times[lti0 - 1]
i["lpc-coefs"].zero_starttime()
i["dur"] = dur
#For windowfactor=2 (save only samples and assume 16kHz)
i["residuals"] = restrack.slice(
restrack.index_at(lpctrack.times[lti0] - pitchperiod[lti0]),
restrack.index_at(lpctrack.times[lti1] + pitchperiod[lti0])).values
return u
def hts_synth(self, utt, processname):
htsparms = self.engine_parms.copy()
htsparms["-of"] = "%(tempolf0_file)s"
if "htsparms" in utt:
htsparms.update(utt["htsparms"]) #parm overrides for this utt...
#build command string and execute:
cmds = self.hts_bin
for k in htsparms:
if htsparms[k]:
if htsparms[k] is True:
cmds += " " + k
else:
cmds += " " + k + " " + str(htsparms[k])
cmds += " %(tempilab_file)s"
fd1, tempwav_file = mkstemp(prefix="ttslab_", suffix=".wav")
fd2, tempilab_file = mkstemp(prefix="ttslab_")
fd3, tempolab_file = mkstemp(prefix="ttslab_")
fd4, tempolf0_file = mkstemp(prefix="ttslab_")
cmds = cmds % {'models_dir': self.models_dir,
'tempwav_file': tempwav_file,
'tempilab_file': tempilab_file,
'tempolab_file': tempolab_file,
'tempolf0_file': tempolf0_file}
#print(cmds)
with codecs.open(tempilab_file, "w", encoding="utf-8") as outfh:
outfh.write("\n".join(utt["hts_label"]))
os.system(cmds)
#load seg endtimes into utt:
with open(tempolab_file) as infh:
lines = infh.readlines()
segs = utt.get_relation("Segment").as_list()
assert len(segs) == len(lines)
for line, seg in zip(lines, segs):
seg["end"] = hts_labels_tone.htk_int_to_float(line.split()[1])
#load audio:
utt["waveform"] = Waveform(tempwav_file)
#load lf0:
f0 = np.exp(np.fromfile(tempolf0_file, "float32")) #load and lf0 to hertz
#to semitones relative to 1Hz:
f0[f0.nonzero()] = 12.0 * np.log2(f0[f0.nonzero()]) # 12 * log2 (F0 / F0reference) where F0reference = 1
f0t = Track()
f0t.values = f0
f0t.times = np.arange(len(f0), dtype=np.float64) * 0.005
utt["f0"] = f0t
#cleanup tempfiles:
os.close(fd1)
os.close(fd2)
os.close(fd3)
os.close(fd4)
os.remove(tempwav_file)
os.remove(tempolab_file)
os.remove(tempilab_file)
os.remove(tempolf0_file)
return utt
