def test_single(trainer, speaker2id_path, result_dir, enc_only, s_speaker,
t_speaker):
with open(speaker2id_path, 'r') as f_json:
speaker2id = json.load(f_json)
if s_speaker == 'S015':
filename = './data/english/train/unit/S015_0361841101.wav'
elif s_speaker == 'S119':
filename = './data/english/train/unit/S119_1561145062.wav'
else:
raise NotImplementedError('Please modify path manually!')
_, spec = get_spectrograms(filename)
spec_expand = np.expand_dims(spec, axis=0)
spec_tensor = torch.from_numpy(spec_expand).type(torch.FloatTensor)
c = torch.tensor(torch.from_numpy(np.array([speaker2id[t_speaker]
]))).cuda()
result = trainer.test_step(spec_tensor, c, enc_only=enc_only)
result = result.squeeze(axis=0).transpose((1, 0))
wav_data = spectrogram2wav(result)
write(os.path.join(result_dir, 'result.wav'), rate=16000, data=wav_data)
print(
'Testing on source speaker {} and target speaker {}, output shape: {}'.
format(s_speaker, t_speaker, result.shape))
def target_classify(trainer, seg_len, synthesis_list, result_dir, flag='test'):
dir_path = os.path.join(result_dir, f'{flag}/')
with open(synthesis_list, 'r') as f:
file = f.readlines()
acc = []
for line in file:
# get wav path
line = line.split('\n')[0].split(' ')
utt_id = line[0].split('/')[1].split('_')[1]
tar_speaker = line[1]
wav_path = os.path.join(dir_path, f'{tar_speaker}_{utt_id}.wav')
# get spectrogram
_, spec = get_spectrograms(wav_path)
# padding spec
if len(spec) < seg_len:
padding = np.zeros((seg_len - spec.shape[0], spec.shape[1]))
spec = np.concatenate((spec, padding), axis=0)
# classification
logits = []
for idx in range(0, len(spec), seg_len):
if idx + (seg_len * 2) > len(spec):
spec_frag = spec[idx:-1]
else:
spec_frag = spec[idx:idx + seg_len]
if len(spec_frag) >= seg_len:
x = torch.from_numpy(
np.expand_dims(spec_frag[:seg_len, :],
axis=0)).type(torch.FloatTensor)
logit = trainer.classify(x)
logits.append(logit)
elif idx == 0:
raise RuntimeError('Please check if input is too short!')
logits = np.concatenate(logits, axis=0)
#logits = np.sum(logits, axis = 0)
for logit in logits:
am = logit.argmax()
if am == 0:
clf_speaker = 'V001'
elif am == 1:
clf_speaker = 'V002'
else:
clf_speaker = 'None'
if clf_speaker == tar_speaker:
acc.append(1)
#print('[info]: {} is classified to {}'.format(wav_path, clf_speaker))
else:
acc.append(0)
#print('[Error]: {} is classified to {}'.format(wav_path, clf_speaker))
print('Classification Acc: {:.3f}'.format(np.sum(acc) / float(len(acc))))
def encode_for_tacotron(target, trainer, seg_len, multi2idx_path, wav_path,
result_path):
wavs = sorted(glob.glob(os.path.join(wav_path, '*.wav')))
print('[Converter] - Number of wav files to encoded: ', len(wavs))
names = []
enc_outputs = []
for wav_path in tqdm(wavs):
name = wav_path.split('/')[-1].split('.')[0]
s_id = name.split('_')[0]
u_id = name.split('_')[1]
if s_id != target:
continue
y, sr = librosa.load(wav_path)
d = librosa.get_duration(y=y, sr=sr)
if d > 25:
continue # --> this filter out too long utts, 3523/3533 for V001 and V002 together in the english dataset
_, spec = get_spectrograms(wav_path)
encodings = encode(spec, trainer, seg_len, save=False)
encodings = parse_encodings(encodings)
enc_outputs.append(encodings)
names.append((s_id, u_id))
# build encodings to character mapping
idx = 0
multi2idx = {}
print('[Converter] - Building encoding to symbol mapping...')
for encodings in tqdm(enc_outputs):
for encoding in encodings:
if str(encoding) not in multi2idx:
multi2idx[str(encoding)] = symbols[idx]
idx += 1
print('[Converter] - Number of unique discret units: ', len(multi2idx))
with open(multi2idx_path, 'w') as file:
file.write(json.dumps(multi2idx))
result_path = result_path.replace('target', target)
print('[Converter] - Writing to meta file...')
with open(result_path, 'w') as file:
for i, encodings in enumerate(enc_outputs):
file.write(str(names[i][0]) + '_' + str(names[i][1] + '|'))
for encoding in encodings:
file.write(multi2idx[str(encoding)])
file.write('\n')
def test_single(trainer, seg_len, speaker2id_path, result_dir, enc_only,
s_speaker, t_speaker):
with open(speaker2id_path, 'r') as f_json:
speaker2id = json.load(f_json)
if s_speaker == 'S015':
filename = './data/english/train/unit/S015_0361841101.wav'
elif s_speaker == 'S119':
filename = './data/english/train/unit/S119_1561145062.wav'
elif s_speaker == 'S130':
filename = './data/english/test/S130_3516588097.wav'
elif s_speaker == 'S089':
filename = './data/english/test/S089_1810826781.wav'
elif s_speaker == 'S378':
filename = './data/surprise/test/S378_117437.wav'
else:
raise NotImplementedError('Please modify path manually!')
_, spec = get_spectrograms(filename)
wav_data, encodings = convert(trainer,
seg_len,
src_speaker_spec=spec,
src_speaker=s_speaker,
tar_speaker=t_speaker,
utt_id='',
speaker2id=speaker2id,
result_dir=result_dir,
enc_only=enc_only,
save=[])
sf.write(os.path.join(result_dir, 'result.wav'), wav_data, hp.sr, 'PCM_16')
write_encodings(os.path.join(result_dir, 'result.txt'), encodings)
err_result = compare_asr(filename, os.path.join(result_dir, 'result.wav'))
print(
'Testing on source speaker {} and target speaker {}, output shape: {}'.
format(s_speaker, t_speaker, wav_data.shape))
print('Comparing ASR result - WERR: {:.3f} CERR: {:.3f}'.format(
err_result[0], err_result[1]))
def main():
#---initialize---#
args = get_test_args()
HPS = Hps(args.hps_path)
hps = HPS.get_tuple()
trainer = get_trainer(args.hps_path, args.encoder_path, hps.g_mode,
hps.enc_mode)
if args.eval_t == 'None':
print(
'[Tacotron] - None is not a valid evaluation target! Please specify target manually, must be either V001, or V002.'
)
return
# Tacotron implementation: https://github.com/andi611/TTS-Tacotron-Pytorch
model = Tacotron(n_vocab=len(symbols),
embedding_dim=config.embedding_dim,
mel_dim=config.num_mels,
linear_dim=config.num_freq,
r=config.outputs_per_step,
padding_idx=config.padding_idx,
attention=config.attention,
use_mask=config.use_mask)
#---handle path---#
result_dir = os.path.join(args.result_dir, args.sub_result_dir)
os.makedirs(result_dir, exist_ok=True)
checkpoint_path = os.path.join(args.ckpt_dir, args.model_name)
if args.dataset == 'english' and not os.path.isdir(
'./ckpt_tacotron_english'):
print(
'[Tacotron] - Recommand using the following name for ckpt_dir: ./ckpt_tacotron_english/'
)
elif args.dataset == 'surprise' and not os.path.isdir(
'./ckpt_tacotron_surprise'):
print(
'[Tacotron] - Recommand using the following name for ckpt_dir: ./ckpt_tacotron_surprise/'
)
#---load and set model---#
print('[Tacotron] - Testing on the {} set.'.format(args.dataset))
print('[Tacotron] - Loading model: ', checkpoint_path)
checkpoint = torch.load(checkpoint_path)
model.load_state_dict(checkpoint["state_dict"])
#---load and set mappings---#
print('[Tacotron] - Loading mapping files: ', args.speaker2id_path)
valid_arguments(valid_target=args.dataset, arg=args.speaker2id_path)
with open(args.speaker2id_path, 'r') as f_json:
speaker2id = json.load(f_json)
print('[Tacotron] - Loading mapping files: ', args.multi2idx_path)
with open(args.multi2idx_path, 'r') as f_json:
multi2idx = json.load(f_json)
if not args.test_single:
#---parse testing list---#
print('[Tacotron] - Testing from list: ', args.synthesis_list)
valid_arguments(valid_target=args.dataset, arg=args.synthesis_list)
feeds = []
with open(args.synthesis_list, 'r') as f:
file = f.readlines()
for line in file:
line = line.split('\n')[0].split(' ')
feeds.append({
's_id': line[0].split('/')[1].split('_')[0],
'utt_id': line[0].split('/')[1].split('_')[1],
't_id': line[1],
})
print('[Tester] - Number of files to be resynthesize: ', len(feeds))
for feed in tqdm(feeds):
if feed['t_id'] == args.eval_t:
wav_path = os.path.join(
args.testing_dir,
feed['s_id'] + '_' + feed['utt_id'] + '.wav')
_, spec = get_spectrograms(wav_path)
encodings = encode(spec, trainer, hps.seg_len, save=False)
encodings = parse_encodings(encodings)
line = ''.join([multi2idx[encoding] for encoding in encodings])
print(line)
out_path = os.path.join(
result_dir, feed['t_id'] + '_' + feed['utt_id'] + '.wav')
synthesis_speech(model, text=line, path=out_path)
else:
wav_path = './data/english/train/voice/V002_0674932509.wav'
# wav_path = './data/english/train/voice/V002_2252538703.wav'
# wav_path = './data/english/train/voice/V002_1665800749.wav'
_, spec = get_spectrograms(wav_path)
encodings = encode(spec, trainer, hps.seg_len, save=False)
write_encodings(path='./result/result.wav', encodings=encodings)
parsed_encodings = parse_encodings(encodings)
line = ''.join([multi2idx[encoding] for encoding in parsed_encodings])
print(line)
synthesis_speech(model, text=line, path='./result/result.wav')
# model.decoder.max_decoder_steps = config.max_decoder_steps # Set large max_decoder steps to handle long sentence outputs
sys.exit(0)