def predict(hparams, model_dir, checkpoint_path, output_dir, test_files):
audio = Audio(hparams)
def predict_input_fn():
records = tf.data.TFRecordDataset(list(test_files))
dataset = DatasetSource(records, hparams)
batched = dataset.make_source_and_target().group_by_batch(
batch_size=1).arrange_for_prediction()
return batched.dataset
estimator = WaveNetModel(hparams, model_dir)
predictions = map(
lambda p: PredictedAudio(p["id"], p["key"], p["predicted_waveform"], p[
"ground_truth_waveform"], p["mel"], p["text"]),
estimator.predict(predict_input_fn, checkpoint_path=checkpoint_path))
for v in predictions:
key = v.key.decode('utf-8')
audio_filename = f"{key}.wav"
audio_filepath = os.path.join(output_dir, audio_filename)
tf.logging.info(f"Saving {audio_filepath}")
audio.save_wav(v.predicted_waveform, audio_filepath)
png_filename = f"{key}.png"
png_filepath = os.path.join(output_dir, png_filename)
tf.logging.info(f"Saving {png_filepath}")
# ToDo: pass global step
plot_wav(png_filepath, v.predicted_waveform, v.ground_truth_waveform,
key, 0, v.text.decode('utf-8'), hparams.sample_rate)
class LJSpeech:
def __init__(self, in_dir, mel_out_dir, wav_out_dir, hparams):
self.in_dir = in_dir
self.mel_out_dir = mel_out_dir
self.wav_out_dir = wav_out_dir
self.audio = Audio(hparams)
@property
def record_ids(self):
return map(lambda v: str(v), range(1, 13101))
def record_file_path(self, record_id, kind):
assert kind in ["source", "target"]
return os.path.join(self.mel_out_dir,
f"ljspeech-{kind}-{int(record_id):05d}.tfrecord")
def text_and_path_rdd(self, sc: SparkContext):
return sc.parallelize(self._extract_all_text_and_path())
def process_wav(self, rdd: RDD):
return rdd.mapValues(self._process_wav)
def _extract_text_and_path(self, line, index):
parts = line.strip().split('|')
key = parts[0]
text = parts[2]
wav_path = os.path.join(self.in_dir, 'wavs', '%s.wav' % key)
return TextAndPath(index, key, wav_path, None, text)
def _extract_all_text_and_path(self):
with open(os.path.join(self.in_dir, 'metadata.csv'),
mode='r',
encoding='utf-8') as f:
for index, line in enumerate(f):
extracted = self._extract_text_and_path(line, index)
if extracted is not None:
yield (index, extracted)
def _process_wav(self, paths: TextAndPath):
wav = self.audio.load_wav(paths.wav_path)
mel_spectrogram = self.audio.melspectrogram(wav).astype(np.float32).T
mel_spectrogram = self.audio.normalize_mel(mel_spectrogram)
mel_filepath = os.path.join(self.mel_out_dir, f"{paths.key}.mfbsp")
wav_filepath = os.path.join(self.wav_out_dir, f"{paths.key}.wav")
mel_spectrogram.tofile(mel_filepath, format="<f4")
self.audio.save_wav(wav, wav_filepath)
def predict(hparams,
model_dir, checkpoint_path, output_dir,
test_source_files, test_target_files):
if hparams.half_precision:
backend.set_floatx(tf.float16.name)
backend.set_epsilon(1e-4)
audio = Audio(hparams)
def predict_input_fn():
source = tf.data.TFRecordDataset(list(test_source_files))
target = tf.data.TFRecordDataset(list(test_target_files))
dataset = dataset_factory(source, target, hparams)
batched = dataset.prepare_and_zip().group_by_batch(
batch_size=1).move_mel_to_source()
return batched.dataset
estimator = model_factory(hparams, model_dir, None)
predictions = map(
lambda p: PredictedMel(p["id"], p["key"], p["mel"], p.get("mel_postnet"), p["mel"].shape[1], p["mel"].shape[0],
p["ground_truth_mel"],
p["alignment"], p.get("alignment2"), p.get("alignment3"), p.get("alignment4"),
p.get("alignment5"), p.get("alignment6"), p.get("attention2_gate_activation"),
p["source"], p["text"], p.get("accent_type"), p),
estimator.predict(predict_input_fn, checkpoint_path=checkpoint_path))
for v in predictions:
key = v.key.decode('utf-8')
mel_filename = f"{key}.{hparams.predicted_mel_extension}"
mel_filepath = os.path.join(output_dir, mel_filename)
ground_truth_mel = v.ground_truth_mel.astype(np.float32)
predicted_mel = v.predicted_mel.astype(np.float32)
mel_denormalized = audio.denormalize_mel(predicted_mel)
linear_spec = audio.logmelspc_to_linearspc(mel_denormalized)
wav = audio.griffin_lim(linear_spec)
audio.save_wav(wav, os.path.join(output_dir, f"{key}.wav"))
assert mel_denormalized.shape[1] == hparams.num_mels
mel_denormalized.tofile(mel_filepath, format='<f4')
text = v.text.decode("utf-8")
plot_filename = f"{key}.png"
plot_filepath = os.path.join(output_dir, plot_filename)
alignments = [x.astype(np.float32) for x in [
v.alignment, v.alignment2, v.alignment3, v.alignment4, v.alignment5, v.alignment6] if x is not None]
if hparams.model == "SSNTModel":
ssnt_metrics.save_alignment_and_log_probs([v.alignment],
[v.all_fields["log_emit_and_shift_probs"]],
[v.all_fields["log_output_prob"]],
[None],
text, v.key, 0,
os.path.join(output_dir, f"{key}_probs.png"))
ssnt_metrics.write_prediction_result(v.id, key, text,
v.all_fields["log_emit_and_shift_probs"],
v.all_fields["log_output_prob"],
os.path.join(output_dir, f"{key}_probs.tfrecord"))
plot_predictions(alignments, ground_truth_mel, predicted_mel, text, v.key, plot_filepath)
prediction_filename = f"{key}.tfrecord"
prediction_filepath = os.path.join(output_dir, prediction_filename)
write_prediction_result(v.id, key, alignments, mel_denormalized, audio.denormalize_mel(ground_truth_mel),
text, v.source, prediction_filepath)
class Synthesizer():
def __init__(self, model_path, out_dir, text_file, sil_file,
use_griffin_lim, hparams):
self.model_path = model_path
self.out_dir = out_dir
self.text_file = text_file
self.sil_file = sil_file
self.use_griffin_lim = use_griffin_lim
self.hparams = hparams
self.model = get_model(model_path, hparams)
self.audio_class = Audio(hparams)
if hparams.use_phone:
from text.phones import Phones
phone_class = Phones(hparams.phone_set_file)
self.text_to_sequence = phone_class.text_to_sequence
else:
from text import text_to_sequence
self.text_to_sequence = text_to_sequence
# self.out_png_dir = os.path.join(self.out_dir, 'png')
# os.makedirs(self.out_png_dir, exist_ok=True)
self.out_wav_dir = os.path.join(self.out_dir, 'wav')
os.makedirs(self.out_wav_dir, exist_ok=True)
def get_inputs(self, meta_data):
hparams = self.hparams
SEQUENCE_ = []
SPEAKERID_ = []
STYLE_ID_ = []
FILENAME_ = []
# Prepare text input
for i in range(len(meta_data)):
filename = meta_data[i].strip().split('|')[1]
print('Filename=', filename)
phone_text = meta_data[i].strip().split('|')[-1]
print('Text=', phone_text)
speaker_id = int(meta_data[i].strip().split('|')[-2])
print('SpeakerID=', speaker_id)
sequence = np.array(
self.text_to_sequence(meta_data[i].strip().split('|')[-1],
['english_cleaners'])) # [None, :]
print(sequence)
sequence = torch.autograd.Variable(
torch.from_numpy(sequence)).to(device).long()
speaker_id = torch.LongTensor(
[speaker_id]).to(device) if hparams.is_multi_speakers else None
style_id = torch.LongTensor([
int(meta_data[i].strip().split('|')[1])
]).to(device) if hparams.is_multi_styles else None
SEQUENCE_.append(sequence)
SPEAKERID_.append(speaker_id)
STYLE_ID_.append(style_id)
FILENAME_.append(filename)
return SEQUENCE_, SPEAKERID_, STYLE_ID_, FILENAME_
def gen_mel(self, meta_data):
SEQUENCE_, SPEAKERID_, STYLE_ID_, FILENAME_ = self.get_inputs(
meta_data)
MEL_OUTOUTS_ = []
FILENAME_NEW_ = []
# Decode text input and plot results
with torch.no_grad():
for i in range(len(SEQUENCE_)):
mel_outputs, _, _ = self.model.inference(
text=SEQUENCE_[i],
spk_ids=SPEAKERID_[i],
utt_mels=TextMelLoader_refine(hparams.training_files,
hparams).utt_mels)
MEL_OUTOUTS_.append(
mel_outputs.transpose(
0, 1).float().data.cpu().numpy()) # (dim, length)
FILENAME_NEW_.append('spkid_' + str(SPEAKERID_[i].item()) +
'_filenum_' + FILENAME_[i])
print('mel_outputs.shape=', mel_outputs.shape)
# image_path = os.path.join(self.out_png_dir, "{}_spec_stop.png".format(filename))
# fig, axes = plt.subplots(1, 1, figsize=(8,8))
# axes.imshow(mel_outputs, aspect='auto', origin='bottom', interpolation='none')
# plt.savefig(image_path, format='png')
# plt.close()
return MEL_OUTOUTS_, FILENAME_NEW_
def gen_wav_griffin_lim(self, mel_outputs, filename):
grf_wav = self.audio_class.inv_mel_spectrogram(mel_outputs)
grf_wav = self.audio_class.inv_preemphasize(grf_wav)
wav_path = os.path.join(self.out_wav_dir, "{}-gl.wav".format(filename))
self.audio_class.save_wav(grf_wav, wav_path)
def inference_f(self):
# print(meta_data['n'])
meta_data = _read_meta_yyh(self.text_file)
MEL_OUTOUTS_, FILENAME_NEW_ = self.gen_mel(meta_data)
for i in range(len(MEL_OUTOUTS_)):
np.save(
os.path.join(
self.out_wav_dir,
"{}.npy".format(FILENAME_NEW_[i] + '_' +
self.model_path.split('/')[-1])),
MEL_OUTOUTS_[i].transpose(1, 0))
if self.use_griffin_lim:
self.gen_wav_griffin_lim(
MEL_OUTOUTS_[i],
FILENAME_NEW_[i] + '_' + self.model_path.split('/')[-1])
source_dir = self.out_wav_dir
target_dir_npy = r'../../../Melgan_pipeline/melgan_file/gen_npy'
target_dir_wav = r'../../../Melgan_pipeline/melgan_file/gen_wav'
target_dir_wav_16k = r'../../../Melgan_pipeline/melgan_file/gen_wav_16k'
for file in sorted(os.listdir(source_dir)):
temp_file_dir = source_dir + os.sep + file
shutil.copy(temp_file_dir, target_dir_npy)
os.remove(temp_file_dir)
os.system(
'cd ../../../Melgan_pipeline/melgan_file/scripts/ && python generate_from_folder_mels_4-4_npy.py --load_path=../../personalvoice/melgan16k --folder=../gen_npy --save_path=../gen_wav'
)
os.system(
'cd ../../../Tools/16kConverter/ && python Convert16k.py --wave_path=../../Melgan_pipeline/melgan_file/gen_wav --output_dir=../../Melgan_pipeline/melgan_file/gen_wav_16k'
)
for file in sorted(os.listdir(target_dir_npy)):
temp_file_dir = target_dir_npy + os.sep + file
os.remove(temp_file_dir)
for file in sorted(os.listdir(target_dir_wav)):
temp_file_dir = target_dir_wav + os.sep + file
os.remove(temp_file_dir)
for file in sorted(os.listdir(target_dir_wav_16k)):
temp_file_dir = target_dir_wav_16k + os.sep + file
shutil.copy(temp_file_dir, source_dir)
os.remove(temp_file_dir)
return print('finished')
class Synthesizer():
def __init__(self, model_path, out_dir, text_file, sil_file,
use_griffin_lim, gen_wavenet_fea, hparams):
self.out_dir = out_dir
self.text_file = text_file
self.sil_file = sil_file
self.use_griffin_lim = use_griffin_lim
self.gen_wavenet_fea = gen_wavenet_fea
self.hparams = hparams
self.model = get_model(model_path, hparams)
self.audio_class = Audio(hparams)
if hparams.use_phone:
from text.phones import Phones
phone_class = Phones(hparams.phone_set_file)
self.text_to_sequence = phone_class.text_to_sequence
else:
from text import text_to_sequence
self.text_to_sequence = text_to_sequence
if hparams.is_multi_speakers and not hparams.use_pretrained_spkemb:
self.speaker_id_dict = gen_speaker_id_dict(hparams)
self.out_png_dir = os.path.join(self.out_dir, 'png')
os.makedirs(self.out_png_dir, exist_ok=True)
if self.use_griffin_lim:
self.out_wav_dir = os.path.join(self.out_dir, 'wav')
os.makedirs(self.out_wav_dir, exist_ok=True)
if self.gen_wavenet_fea:
self.out_mel_dir = os.path.join(self.out_dir, 'mel')
os.makedirs(self.out_mel_dir, exist_ok=True)
def get_mel_gt(self, wavname):
hparams = self.hparams
if not hparams.load_mel:
if hparams.use_hdf5:
with h5py.File(hparams.hdf5_file, 'r') as h5:
data = h5[wavname][:]
else:
filename = os.path.join(hparams.wav_dir, wavname + '.wav')
sr_t, audio = wavread(filename)
assert sr_t == hparams.sample_rate
audio_norm = audio / hparams.max_wav_value
wav = self.audio_class._preemphasize(audio_norm)
melspec = self.audio_class.melspectrogram(wav, clip_norm=True)
melspec = torch.FloatTensor(melspec.astype(np.float32))
else:
if hparams.use_zip:
with zipfile.ZipFile(hparams.zip_path, 'r') as f:
data = f.read(wavname)
melspec = np.load(io.BytesIO(data))
melspec = torch.FloatTensor(melspec.astype(np.float32))
elif hparams.use_hdf5:
with h5py.File(hparams.hdf5_file, 'r') as h5:
melspec = h5[wavname][:]
melspec = torch.FloatTensor(melspec.astype(np.float32))
else:
filename = os.path.join(hparams.wav_dir, wavname + '.npy')
melspec = torch.from_numpy(np.load(filename))
melspec = torch.unsqueeze(melspec, 0)
return melspec
def get_inputs(self, meta_data):
hparams = self.hparams
# Prepare text input
# filename = meta_data['n']
# filename = os.path.splitext(os.path.basename(filename))[0]
filename = meta_data[0].strip().split('|')[0]
print(meta_data[0].strip().split('|')[-1])
print(meta_data[0].strip().split('|')[1])
sequence = np.array(
self.text_to_sequence(meta_data[0].strip().split('|')[-1],
['english_cleaners'])) # [None, :]
# sequence = torch.autograd.Variable(
# torch.from_numpy(sequence)).cuda().long()
print(sequence)
sequence = torch.autograd.Variable(
torch.from_numpy(sequence)).to(device).long()
if hparams.is_multi_speakers:
if hparams.use_pretrained_spkemb:
ref_file = meta_data['r']
spk_embedding = np.array(np.load(ref_file))
spk_embedding = torch.autograd.Variable(
torch.from_numpy(spk_embedding)).to(device).float()
inputs = (sequence, spk_embedding)
else:
speaker_name = filename.split('_')[0]
speaker_id = self.speaker_id_dict[speaker_name]
speaker_id = np.array([speaker_id])
# speaker_id = torch.autograd.Variable(
# torch.from_numpy(speaker_id)).cuda().long()
speaker_id = torch.autograd.Variable(
torch.from_numpy(speaker_id)).to(device).long()
inputs = (sequence, speaker_id)
if hparams.is_multi_styles:
style_id = np.array([int(meta_data[0].strip().split('|')[1])])
style_id = torch.autograd.Variable(
torch.from_numpy(style_id)).to(device).long()
inputs = (sequence, style_id)
elif hparams.use_vqvae:
ref_file = meta_data['r']
spk_ref = self.get_mel_gt(ref_file)
inputs = (sequence, spk_ref)
else:
inputs = (sequence)
return inputs, filename
def gen_mel(self, meta_data):
inputs, filename = self.get_inputs(meta_data)
speaker_id = None
style_id = None
spk_embedding = None
spk_ref = None
if self.hparams.is_multi_speakers:
if self.hparams.use_pretrained_spkemb:
sequence, spk_embedding = inputs
else:
sequence, speaker_id = inputs
elif hparams.use_vqvae:
sequence, spk_ref = inputs
else:
sequence = inputs
if self.hparams.is_multi_styles:
sequence, style_id = inputs
# Decode text input and plot results
with torch.no_grad():
mel_outputs, gate_outputs, att_ws = self.model.inference(
sequence,
self.hparams,
spk_id=speaker_id,
style_id=style_id,
spemb=spk_embedding,
spk_ref=spk_ref)
duration_list = DurationCalculator._calculate_duration(att_ws)
print('att_ws.shape=', att_ws.shape)
print('duration=', duration_list)
print('duration_sum=', torch.sum(duration_list))
print('focus_rete=',
DurationCalculator._calculate_focus_rete(att_ws))
# print(mel_outputs.shape) # (length, dim)
mel_outputs = mel_outputs.transpose(
0, 1).float().data.cpu().numpy() # (dim, length)
mel_outputs_with_duration = get_duration_matrix(
char_text_dir=self.text_file,
duration_tensor=duration_list,
save_mode='phone').transpose(0, 1).float().data.cpu().numpy()
gate_outputs = gate_outputs.float().data.cpu().numpy()
att_ws = att_ws.float().data.cpu().numpy()
image_path = os.path.join(self.out_png_dir,
"{}_att.png".format(filename))
_plot_and_save(att_ws, image_path)
image_path = os.path.join(self.out_png_dir,
"{}_spec_stop.png".format(filename))
fig, axes = plt.subplots(3, 1, figsize=(8, 8))
axes[0].imshow(mel_outputs,
aspect='auto',
origin='bottom',
interpolation='none')
axes[1].imshow(mel_outputs_with_duration,
aspect='auto',
origin='bottom',
interpolation='none')
axes[2].scatter(range(len(gate_outputs)),
gate_outputs,
alpha=0.5,
color='red',
marker='.',
s=5,
label='predicted')
plt.savefig(image_path, format='png')
plt.close()
return mel_outputs, filename
def gen_wav_griffin_lim(self, mel_outputs, filename):
grf_wav = self.audio_class.inv_mel_spectrogram(mel_outputs)
grf_wav = self.audio_class.inv_preemphasize(grf_wav)
wav_path = os.path.join(self.out_wav_dir, "{}-gl.wav".format(filename))
self.audio_class.save_wav(grf_wav, wav_path)
def gen_wavenet_feature(self, mel_outputs, filename, add_end_sil=True):
# denormalize
mel = self.audio_class._denormalize(mel_outputs)
# normalize to 0-1
mel = np.clip(((mel - self.audio_class.hparams.min_level_db) /
(-self.audio_class.hparams.min_level_db)), 0, 1)
mel = mel.T.astype(np.float32)
frame_size = 200
SILSEG = 0.3
SAMPLING = 16000
sil_samples = int(SILSEG * SAMPLING)
sil_frames = int(sil_samples / frame_size)
sil_data, _ = soundfile.read(self.sil_file)
sil_data = sil_data[:sil_samples]
sil_mel_spec, _ = self.audio_class._magnitude_spectrogram(
sil_data, clip_norm=True)
sil_mel_spec = (sil_mel_spec + 4.0) / 8.0
pad_mel_data = np.concatenate((sil_mel_spec[:sil_frames], mel), axis=0)
if add_end_sil:
pad_mel_data = np.concatenate(
(pad_mel_data, sil_mel_spec[:sil_frames]), axis=0)
out_mel_file = os.path.join(self.out_mel_dir,
'{}-wn.mel'.format(filename))
save_htk_data(pad_mel_data, out_mel_file)
def inference_f(self):
# print(meta_data['n'])
meta_data = _read_meta_yyh(self.text_file)
mel_outputs, filename = self.gen_mel(meta_data)
print('my_mel_outputs=', mel_outputs)
print('my_mel_outputs_max=', np.max(mel_outputs))
print('my_mel_outputs_min=', np.min(mel_outputs))
mel_outputs = np.load(r'../out_0.npy').transpose(1, 0)
mel_outputs = mel_outputs * 8.0 - 4.0
print('his_mel_outputs=', mel_outputs)
print('his_mel_outputs_max=', np.max(mel_outputs))
print('his_mel_outputs_min=', np.min(mel_outputs))
if self.use_griffin_lim:
self.gen_wav_griffin_lim(mel_outputs, filename)
if self.gen_wavenet_fea:
self.gen_wavenet_feature(mel_outputs, filename)
return filename
def inference(self):
all_meta_data = _read_meta(self.text_file, hparams.meta_format)
list(map(self.inference_f, all_meta_data))