def gen_embedding(speaker):
training_list = hparams.training_list
train_set_A = TextMelIDLoader(training_list,
hparams.mel_mean_std,
hparams.speaker_A,
hparams.speaker_B,
shuffle=False,
pids=[speaker])
collate_fn = TextMelIDCollate(
lcm(hparams.n_frames_per_step_encoder,
hparams.n_frames_per_step_decoder))
train_loader_A = DataLoader(train_set_A,
num_workers=1,
shuffle=False,
sampler=None,
batch_size=1,
pin_memory=False,
drop_last=True,
collate_fn=collate_fn)
with torch.no_grad():
speaker_embeddings = []
for i, batch in enumerate(train_loader_A):
#print i
x, y = model.parse_batch(batch)
text_input_padded, mel_padded, text_lengths, mel_lengths, speaker_id = x
speaker_id, speaker_embedding = model.speaker_encoder.inference(
mel_padded)
speaker_embedding = speaker_embedding.data.cpu().numpy()
#speaker_embedding = speaker_embedding * 0.1
speaker_embeddings.append(speaker_embedding)
speaker_embeddings = np.vstack(speaker_embeddings)
print(speaker_embeddings.shape)
if not os.path.exists('outdir/embeddings'):
os.makedirs('outdir/embeddings')
np.save('outdir/embeddings/%s.npy' % speaker, speaker_embeddings)
plot_data([speaker_embeddings], 'outdir/embeddings/%s.pdf' % speaker)
def generate(loader,
reference_mel,
beam_width,
path_save,
ref_sp,
sample_list,
num=10,
input_text=False):
with torch.no_grad():
errs = []
totalphs = []
for i, batch in enumerate(loader):
if i == num:
break
#sample_id = sample_list[i].split('/')[-1][9:17+4]
sample_id = sample_list[i].split('/')[-1]
print(('index:%d, decoding %s ...' % (i, sample_id)))
text_file = '{}.txt'.format(os.path.basename(sample_list[i]))
text_path = os.path.dirname(sample_list[i]).replace(
'spec-wgan', 'text')
text_path = os.path.join(text_path, text_file)
text = open(text_path, 'r').readlines()[0].rstrip()
print('{}: {}'.format(text_file, text))
text_path_output = os.path.join(path_save,
'txt/Txt_{}'.format(text_file))
copyfile(text_path, text_path_output)
wav_file = '{}.wav'.format(os.path.basename(sample_list[i]))
wav_path = os.path.dirname(sample_list[i]).replace(
'spec-wgan', 'wav22_silence_trimmed')
wav_path = os.path.join(wav_path, wav_file)
wav_path_output = os.path.join(path_save,
'wav/Wav_{}'.format(wav_file))
copyfile(wav_path, wav_path_output)
x, y = model.parse_batch(batch)
predicted_mel, post_output, predicted_stop, alignments, \
text_hidden, audio_seq2seq_hidden, audio_seq2seq_phids, audio_seq2seq_alignments, \
speaker_id = model.inference(x, input_text, reference_mel, beam_width)
post_output = post_output.data.cpu().numpy()[0]
alignments = alignments.data.cpu().numpy()[0].T
audio_seq2seq_alignments = audio_seq2seq_alignments.data.cpu(
).numpy()[0].T
text_hidden = text_hidden.data.cpu().numpy(
)[0].T #-> [hidden_dim, max_text_len]
audio_seq2seq_hidden = audio_seq2seq_hidden.data.cpu().numpy()[0].T
audio_seq2seq_phids = audio_seq2seq_phids.data.cpu().numpy()[
0] # [T + 1]
speaker_id = speaker_id.data.cpu().numpy()[0] # scalar
task = 'TTS' if input_text else 'VC'
wav_path = os.path.join(
path_save,
'wav_mel/Wav_%s_ref_%s_%s.wav' % (sample_id, ref_sp, task))
# recover_wav(post_output, wav_path, hparams.mel_mean_std, ismel=ISMEL)
recover_wav_wgan(post_output,
wav_path,
hparams.mel_mean_std,
ismel=ISMEL,
n_fft=1024,
win_length=1024,
hop_length=256)
post_output_path = os.path.join(
path_save,
'mel/Mel_%s_ref_%s_%s.npy' % (sample_id, ref_sp, task))
np.save(post_output_path, post_output.T)
plot_data([alignments, audio_seq2seq_alignments],
os.path.join(
path_save, 'ali/Ali_%s_ref_%s_%s.pdf' %
(sample_id, ref_sp, task)))
plot_data([np.hstack([text_hidden, audio_seq2seq_hidden])],
os.path.join(
path_save, 'hid/Hid_%s_ref_%s_%s.pdf' %
(sample_id, ref_sp, task)))
audio_seq2seq_phids = [
id2ph[id] for id in audio_seq2seq_phids[:-1]
]
target_text = y[0].data.cpu().numpy()[0]
target_text = [id2ph[id] for id in target_text[:]]
if not input_text:
#print 'Sounds like %s, Decoded text is '%(id2sp[speaker_id])
print(audio_seq2seq_phids)
print(target_text)
err = levenshteinDistance(audio_seq2seq_phids, target_text)
print(err, len(target_text))
errs.append(err)
totalphs.append(len(target_text))
# save phone error rate
per = float(sum(errs)) / float(sum(totalphs))
per_file = os.path.join(path_save, 'per.txt')
open(per_file, 'w').writelines('{:.3f}\n'.format(per))
open(per_file, 'a').writelines(','.join([str(e) for e in errs]) + '\n')
open(per_file, 'a').writelines(','.join([str(l) for l in totalphs]) + '\n')
# print float(errs)/float(totalphs)
# return float(errs)/float(totalphs)
return per
def generate(loader, reference_mel, beam_width, path_save, ref_sp,
sample_list, num=10, input_text=False):
with torch.no_grad():
errs = 0
totalphs = 0
for i, batch in enumerate(loader):
if i == num:
break
sample_id = sample_list[i].split('/')[-1][9:17+4]
print(('%d index %s, decoding ...'%(i,sample_id)))
x, y = model.parse_batch(batch)
predicted_mel, post_output, predicted_stop, alignments, \
text_hidden, audio_seq2seq_hidden, audio_seq2seq_phids, audio_seq2seq_alignments, \
speaker_id = model.inference(x, input_text, reference_mel, beam_width)
post_output = post_output.data.cpu().numpy()[0]
alignments = alignments.data.cpu().numpy()[0].T
audio_seq2seq_alignments = audio_seq2seq_alignments.data.cpu().numpy()[0].T
text_hidden = text_hidden.data.cpu().numpy()[0].T #-> [hidden_dim, max_text_len]
audio_seq2seq_hidden = audio_seq2seq_hidden.data.cpu().numpy()[0].T
audio_seq2seq_phids = audio_seq2seq_phids.data.cpu().numpy()[0] # [T + 1]
speaker_id = speaker_id.data.cpu().numpy()[0] # scalar
task = 'TTS' if input_text else 'VC'
recover_wav(post_output,
os.path.join(path_save, 'wav_mel/Wav_%s_ref_%s_%s.wav'%(sample_id, ref_sp, task)),
hparams.mel_mean_std,
ismel=ISMEL)
post_output_path = os.path.join(path_save, 'mel/Mel_%s_ref_%s_%s.npy'%(sample_id, ref_sp, task))
np.save(post_output_path, post_output)
#plot_data([alignments, audio_seq2seq_alignments],
# os.path.join(path_save, 'ali/Ali_%s_ref_%s_%s.pdf'%(sample_id, ref_sp, task)))
plot_data([audio_seq2seq_alignments.T, alignments],
os.path.join(path_save, 'ali/Ali_%s_ref_%s_%s.pdf'%(sample_id, ref_sp, task)))
plot_data([np.hstack([text_hidden, audio_seq2seq_hidden])],
os.path.join(path_save, 'hid/Hid_%s_ref_%s_%s.pdf'%(sample_id, ref_sp, task)))
#plot_data([np.hstack([audio_seq2seq_hidden，text_hidden])],
#os.path.join(path_save, 'hid/Hid_%s_ref_%s_%s.pdf'%(sample_id, ref_sp, task)))
audio_seq2seq_phids = [id2ph[id] for id in audio_seq2seq_phids[:-1]]
target_text = y[0].data.cpu().numpy()[0]
target_text = [id2ph[id] for id in target_text[:]]
if not input_text:
#print 'Sounds like %s, Decoded text is '%(id2sp[speaker_id])
print(audio_seq2seq_phids)
print(target_text)
err = levenshteinDistance(audio_seq2seq_phids, target_text)
print(err, len(target_text))
errs += err
totalphs += len(target_text)
#print float(errs)/float(totalphs)
return float(errs)/float(totalphs)