def create_write_files(ret,sess,g,x,mname,cdir,typeS):
x = np.expand_dims(x, axis=0)
mel_output = np.zeros((1, hp.T_y // hp.r, hp.n_mels * hp.r), np.float32)
decoder_output = np.zeros((1, hp.T_y // hp.r, hp.embed_size), np.float32)
alignments_li = np.zeros((hp.dec_layers, hp.T_x, hp.T_y//hp.r), np.float32)
prev_max_attentions_li = np.zeros((hp.dec_layers, 1), np.int32)
for j in range(hp.T_y // hp.r):
_gs, _mel_output, _decoder_output, _max_attentions_li, _alignments_li = \
sess.run([g.global_step, g.mel_output, g.decoder_output, g.max_attentions_li, g.alignments_li],
{g.x: x,
g.y1: mel_output,
g.prev_max_attentions_li:prev_max_attentions_li})
mel_output[:, j, :] = _mel_output[:, j, :]
decoder_output[:, j, :] = _decoder_output[:, j, :]
prev_max_attentions_li = np.array(_max_attentions_li)[:, :, j]
#mag_output = sess.run([g.mag_output], {g.mel_output: mel_output})
mag_output = sess.run(g.mag_output, {g.decoder_output: decoder_output})
x = np.squeeze(x, axis=0)
txt = invert_text(x)
mag_output = np.squeeze(mag_output[0])
try:
wav = spectrogram2wav(mag_output)
wav, _ = librosa.effects.trim(wav)
write(cdir + "/{}mag.wav".format(mname), hp.sr, wav)
ret.append([txt,wav,typeS+"_world"])
except Exception:
sys.exc_clear()
return ret
def create_write_files(ret, sess, g, x, mname, cdir, typeS):
x = np.expand_dims(x, axis=0)
x = np.append(x, np.zeros((hp.batch_size - 1, hp.T_x)), axis=0)
mel_output = np.zeros((hp.batch_size, hp.T_y // hp.r, hp.n_mels * hp.r),
np.float32)
_gs, mel_output = \
sess.run([g.global_step, g.mel_output],
{g.x: x,
g.y1: mel_output})
mag_output = sess.run(g.mag_output, {g.converter_input: mel_output})
# x = np.squeeze(x[0], axis=0)
x = x[0]
txt = invert_text(x)
mag_output = np.squeeze(mag_output[0])
try:
wav = spectrogram2wav(mag_output)
wav, _ = librosa.effects.trim(wav)
write(cdir + "/{}mag.wav".format(mname), hp.sr, wav)
ret.append([txt, wav, typeS + "_world", mel_output, mag_output])
except Exception:
sys.exc_clear()
return ret
def create_write_files_conv(ret, sess, mel_in, g, x, mname, cdir, typeS):
mag_output = sess.run(g.mag_output, {g.converter_input: mel_in})
txt = invert_text(x)
mag_output = np.squeeze(mag_output[0])
try:
wav = spectrogram2wav(mag_output)
wav, _ = librosa.effects.trim(wav)
write(cdir + "/{}mag.wav".format(mname), hp.sr, wav)
ret.append([txt, wav, typeS + "_world", mel_in, mag_output])
except Exception:
sys.exc_clear()
return ret
def create_write_files(ret, sess, g, x, mname, cdir, samples, typeS):
# for i in range(0, len(x), hp.batch_size):
# x2 = x[i:i+hp.batch_size]
# Get melspectrogram
mel_output = np.zeros((hp.batch_size, hp.T_y // hp.r, hp.n_mels * hp.r),
np.float32)
decoder_output = np.zeros((hp.batch_size, hp.T_y // hp.r, hp.embed_size),
np.float32)
alignments_li = np.zeros((hp.dec_layers, hp.T_x, hp.T_y // hp.r),
np.float32)
prev_max_attentions_li = np.zeros((hp.dec_layers, hp.batch_size), np.int32)
#alignments = np.zeros((hp.T_x, hp.T_y//hp.r), np.float32)
for j in range((hp.T_y // hp.r) // hp.rwin):
_gs, _mel_output, _decoder_output, _max_attentions_li, _alignments_li = \
sess.run([g.global_step, g.mel_output, g.decoder_output, g.max_attentions_li, g.alignments_li],
{g.x: x,
g.y1: mel_output,
g.prev_max_attentions_li:prev_max_attentions_li})
mel_output[:, j * hp.rwin:(j + 1) *
hp.rwin, :] = _mel_output[:,
j * hp.rwin:(j + 1) * hp.rwin, :]
decoder_output[:, j * hp.rwin:(j + 1) *
hp.rwin, :] = _decoder_output[:, j * hp.rwin:(j + 1) *
hp.rwin, :]
#alignments_li[:, :, j*hp.rwin:(j+1)*hp.rwin] = np.array(_alignments_li)[:, :, j*hp.rwin:(j+1)*hp.rwin]
prev_max_attentions_li = np.array(_max_attentions_li)[:, :,
j * hp.rwin]
# Get magnitude
mag_output = sess.run(g.mag_output, {g.decoder_output: decoder_output})
z_list = random.sample(range(0, hp.batch_size), samples)
# Generate wav files
for i, mag in enumerate(mag_output):
if i in z_list:
# generate wav files
#mag = mag*hp.mag_std + hp.mag_mean # denormalize
#audio = spectrogram2wav(np.power(10, mag) ** hp.sharpening_factor)
txt = x[i]
txt = invert_text(txt)
wav = spectrogram2wav(mag)
write(cdir + "/{}_{}.wav".format(mname, i), hp.sr, wav)
ret.append([txt, wav, typeS])
return ret