def gmf_iaif(
psd,
glottal_order=3,
vt_order=48,
):
"""
psd : 1-d numpy array, shape : fft_size/2 +1
"""
fft_size = (psd.shape[0] - 1) * 2
psd_res = psd
for i in range(3):
lpc_coef, g = lpc.psd2lpc(psd_res, order=1)
re_psd = lpc.lpc2psd(lpc_coef, g, fft_size)
psd_res = np.power(np.sqrt(psd_res) / np.sqrt(re_psd), 2)
# vt_order vocal tract
lpc_coef, g = lpc.psd2lpc(psd_res, order=vt_order)
re_psd = lpc.lpc2psd(lpc_coef, g, fft_size)
gross_vocal_psd = re_psd
glottal_psd = psd / gross_vocal_psd
# final glottal_order glottal
glottal_lpc, glottal_g = lpc.psd2lpc(glottal_psd, order=glottal_order)
re_glottal_psd = lpc.lpc2psd(glottal_lpc, glottal_g, fft_size)
vocal_psd = psd / re_glottal_psd
# final vocal tract
vt_lpc, vt_g = lpc.psd2lpc(vocal_psd, order=vt_order)
#re_vt_psd = lpc.lpc2psd(vt_lpc,vt_g,fft_size)
return glottal_lpc, glottal_g, vt_lpc, vt_g
def synthesisLPC2PSD(source_object, filter_object, seeds_signals):
fft_size = (filter_object["spectrogram"].shape[0] - 1) * 2
print("synthesis lpc 2 psd")
excitation_signal, _, _ = synthesisRequiem.get_excitation_signal(
source_object['temporal_positions'], filter_object['fs'],
source_object['f0'], source_object['vuv'], seeds_signals['pulse'],
seeds_signals['noise'], source_object['aperiodicity'])
psd = filter_object['spectrogram']
lpcs = []
gains = []
start_time = time.time()
for i in range(psd.shape[1]):
lpc_coef, g = lpc.psd2lpc(psd[:, i], order=ORDER)
lpcs.append(lpc_coef)
gains.append(g)
print("lpc coef cost {}".format(time.time() - start_time))
recons_psds = []
start_time = time.time()
for lpc_coef, g in zip(lpcs, gains):
recons_psd = lpc.lpc2psd(lpc_coef, g, fft_size)
recons_psds.append(recons_psd)
recons_psds = np.array(recons_psds)
y0 = synthesisRequiem.get_waveform(excitation_signal,
np.transpose(recons_psds, [1, 0]),
source_object['temporal_positions'],
source_object['f0'],
filter_object['fs'])
print("filter cost {}".format(time.time() - start_time))
return y0
def synthesisLPC(source_object, filter_object, seeds_signals):
fft_size = (filter_object["spectrogram"].shape[0] - 1) * 2
print("synthesis lpc")
excitation_signal, _, _ = synthesisRequiem.get_excitation_signal(
source_object['temporal_positions'], filter_object['fs'],
source_object['f0'], source_object['vuv'], seeds_signals['pulse'],
seeds_signals['noise'], source_object['aperiodicity'])
psd = filter_object['spectrogram']
lpcs = []
gains = []
start_time = time.time()
for i in range(psd.shape[1]):
lpc_coef, g = lpc.psd2lpc(psd[:, i], order=ORDER)
lpcs.append(lpc_coef)
gains.append(g)
print("lpc coef cost {}".format(time.time() - start_time))
lpcs = np.array(lpcs)
gains = np.array(gains)
lpcs[:, 0] = np.log(gains)
poledf = pysptk.synthesis.AllPoleDF(ORDER)
synthesizer = pysptk.synthesis.Synthesizer(
poledf, int(filter_object['fs'] * 0.005))
start_time = time.time()
y = synthesizer.synthesis(excitation_signal, lpcs)
print("filter cost {}".format(time.time() - start_time))
return y
def synthesisGMF_IAIF(source_object, filter_object, seeds_signals):
fft_size = (filter_object["spectrogram"].shape[0] - 1) * 2
print("synthesis gmf-iaif")
excitation_signal, _, _ = synthesisRequiem.get_excitation_signal(
source_object['temporal_positions'], filter_object['fs'],
source_object['f0'], source_object['vuv'], seeds_signals['pulse'],
seeds_signals['noise'], source_object['aperiodicity'])
psd = filter_object['spectrogram']
glottal_lpcs = []
glottal_gains = []
lpcs = []
gains = []
start_time = time.time()
for i in range(psd.shape[1]):
lpc_coef, g = lpc.psd2lpc(psd[:, i], order=ORDER)
glottal_lpc, glottal_g, lpc_coef, g = gmf_iaif.gmf_iaif(psd[:, i],
vt_order=ORDER)
glottal_lpcs.append(glottal_lpc)
glottal_gains.append(glottal_g)
lpcs.append(lpc_coef)
gains.append(g)
print("gmf-iaif cost :{}, frames: {}".format(time.time() - start_time,
psd.shape[1]))
recons_psds = []
start_time = time.time()
for glottal_lpc, glottal_g, lpc_coef, g in zip(glottal_lpcs, glottal_gains,
lpcs, gains):
recons_psd = lpc.lpc2psd(glottal_lpc, glottal_g * g, fft_size)
#recons_psd *= lpc.lpc2psd(lpc_coef,g,fft_size)
recons_psds.append(recons_psd)
recons_psds = np.array(recons_psds)
y0 = synthesisRequiem.get_waveform(excitation_signal,
np.transpose(recons_psds, [1, 0]),
source_object['temporal_positions'],
source_object['f0'],
filter_object['fs'])
print("filter cost {}".format(time.time() - start_time))
return y0
def psd2lpc(psd):
glottal_lpcs = []
glottal_gains = []
lpcs = []
gains = []
start_time = time.time()
for i in range(psd.shape[1]):
lpc_coef, g = lpc.psd2lpc(psd[:, i], order=ORDER)
glottal_lpc, glottal_g, lpc_coef, g = gmf_iaif.gmf_iaif(psd[:, i],
vt_order=ORDER)
glottal_lpcs.append(glottal_lpc)
glottal_gains.append(glottal_g)
lpcs.append(lpc_coef)
gains.append(g)
print("gmf-iaif cost :{}, frames: {}".format(time.time() - start_time,
psd.shape[1]))
return np.array(glottal_lpcs), np.array(glottal_gains), np.array(
lpcs), np.array(gains)
