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 synthesisPSD(source_object, filter_object, seeds_signals):
fft_size = (filter_object["spectrogram"].shape[0] - 1) * 2
print("synthesis 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'])
start_time = time.time()
y = synthesisRequiem.get_waveform(excitation_signal,
filter_object['spectrogram'],
source_object['temporal_positions'],
source_object['f0'], filter_object['fs'])
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
recons_vt_psds = np.array(recons_vt_psds)
df = pysptk.synthesis.AllZeroDF(ORDER)
synthesizer = pysptk.synthesis.Synthesizer(df, int(fs * 0.005))
x_glottal_res_zerodf = synthesizer.synthesis(x, lpcs / gains)
df = pysptk.synthesis.AllZeroDF(3)
synthesizer = pysptk.synthesis.Synthesizer(df, int(fs * 0.005))
x_res_zerodf = synthesizer.synthesis(x_glottal_res_zerodf,
glottal_lpcs / glottal_gains)
wavwrite('x_glottal_res_zerodf.wav', fs,
(x_glottal_res_zerodf * 2**15).astype(np.int16))
wavwrite('x_res_zerodf.wav', fs, (x_res_zerodf * 2**15).astype(np.int16))
y = synthesisRequiem.get_waveform(x_res_zerodf,
np.transpose(recons_psds, [1, 0]),
dat['temporal_positions'], dat['f0'],
dat['fs'])
y_from_glottal = synthesisRequiem.get_waveform(
x_glottal_res_zerodf, np.transpose(recons_vt_psds, [1, 0]),
dat['temporal_positions'], dat['f0'], dat['fs'])
wavwrite('x_recons_zerodf.wav', fs, (y * 2**15).astype(np.int16))
wavwrite('x_recons_glottal_zerodf.wav', fs,
(y_from_glottal * 2**15).astype(np.int16))
x_res = np.zeros([x.shape[0] + 100])
x_glottal_res = np.zeros([x.shape[0] + 100])
temporal_positions = dat['temporal_positions']
f0_sequence = dat['f0']
tmp_complex_cepstrum = np.zeros(fft_size)
tmp_complex_cepstrum[latter_index.astype(int) -
1] = tmp_cepstrum[latter_index.astype(int) -
1] * 2
tmp_complex_cepstrum[0] = tmp_cepstrum[0]
spectrum = np.exp(np.fft.ifft(tmp_complex_cepstrum))
response = ifft(1 / spectrum * fft(win_x, fft_size)).real
inv = response
origin = int(temporal_position * fs + 0.501) + 1
safe_index = np.minimum(
len(x_res) - 1, np.arange(origin, origin + fft_size))
x_res[safe_index] += inv
print("filter cost {}".format(time.time() - start_time))
return x_res
x_res = inverse_psd(x, dat)
os.makedirs(out_path, exist_ok=True)
wavwrite(os.path.join(out_path, 'x_res.wav'), fs,
(x_res * 2**15).astype(np.int16))
y = synthesisRequiem.get_waveform(x_res, dat['spectrogram'],
dat['temporal_positions'], dat['f0'],
dat['fs'])
wavwrite(os.path.join(out_path, 'x_recons.wav'), fs,
(y * 2**15).astype(np.int16))