class LJSpeech:
def __init__(self, in_dir, out_dir, hparams):
self.in_dir = in_dir
self.out_dir = out_dir
self.audio = Audio(hparams)
def text_and_audio_path_rdd(self, sc: SparkContext):
return sc.parallelize(self._extract_all_text_and_audio_path())
def process_data(self, rdd: RDD):
return rdd.mapValues(self._process_source_and_target)
def _extract_text_and_path(self, line, index):
parts = line.strip().split('|')
key = parts[0]
wav_path = os.path.join(self.in_dir, 'wavs', '%s.wav' % parts[0])
text = parts[2]
return TextAndAudioPath(index, key, wav_path, text)
def _extract_all_text_and_audio_path(self):
index = 1
with open(os.path.join(self.in_dir, 'metadata.csv'),
mode='r',
encoding='utf-8') as f:
for line in f:
extracted = self._extract_text_and_path(line, index)
if extracted is not None:
yield (index, extracted)
index += 1
def _process_source_and_target(self, paths: TextAndAudioPath):
wav = self.audio.load_wav(paths.wav_path)
n_samples = len(wav)
mel_spectrogram = self.audio.melspectrogram(wav).astype(np.float32).T
n_frames = mel_spectrogram.shape[0]
filename = f"{paths.key}.tfrecord"
filepath = os.path.join(self.out_dir, filename)
tfrecord.write_preprocessed_data(paths.id, paths.key, wav,
mel_spectrogram, paths.text, filepath)
return SourceAndTargetMetaData(paths.id, paths.key, n_samples,
n_frames, filepath)
def _process_mel(self, paths: TextAndAudioPath):
wav = self.audio.load_wav(paths.wav_path)
mel_spectrogram = self.audio.melspectrogram(wav).astype(np.float32).T
sum_mel_powers = np.sum(mel_spectrogram, axis=1)
n_frames = mel_spectrogram.shape[0]
return MelMetaData(n_frames, sum_mel_powers)
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)
class VCTK:
def __init__(self,
in_dir,
out_dir,
hparams,
speaker_info_filename='speaker-info.txt'):
self.in_dir = in_dir
self.out_dir = out_dir
self.speaker_info_filename = speaker_info_filename
self.audio = Audio(hparams)
def list_files(self):
def wav_files(speaker_info: SpeakerInfo):
wav_dir = os.path.join(self.in_dir, f"wav48/p{speaker_info.id}")
return [
os.path.join(wav_dir, wav_file)
for wav_file in sorted(os.listdir(wav_dir))
if wav_file.endswith('.wav')
]
def text_files(speaker_info: SpeakerInfo):
txt_dir = os.path.join(self.in_dir, f"txt/p{speaker_info.id}")
return [
os.path.join(txt_dir, txt_file)
for txt_file in sorted(os.listdir(txt_dir))
if txt_file.endswith('.txt')
]
def text_and_wav_records(file_pairs, speaker_info):
def create_record(txt_f, wav_f, speaker_info):
key1 = os.path.basename(wav_f).strip('.wav')
key2 = os.path.basename(txt_f).strip('.txt')
assert key1 == key2
return TxtWavRecord(0, key1, txt_f, wav_f, speaker_info)
return [
create_record(txt_f, wav_f, speaker_info)
for txt_f, wav_f in file_pairs
]
records = sum([
text_and_wav_records(zip(text_files(si), wav_files(si)), si)
for si in self._load_speaker_info()
], [])
return [
TxtWavRecord(i, r.key, r.txt_path, r.wav_path, r.speaker_info)
for i, r in enumerate(records)
]
def process_sources(self, rdd: RDD):
return rdd.map(self._process_txt)
def process_targets(self, rdd: RDD):
return TargetRDD(
rdd.map(self._process_wav).persist(StorageLevel.MEMORY_AND_DISK))
def _load_speaker_info(self):
with open(os.path.join(self.in_dir, self.speaker_info_filename),
mode='r',
encoding='utf8') as f:
for l in f.readlines()[1:]:
si = l.split()
gender = 0 if si[2] == 'F' else 1
if str(si[0]) != "315": # FixMe: Why 315 is missing?
yield SpeakerInfo(int(si[0]), int(si[1]), gender)
def _process_wav(self, record: TxtWavRecord):
wav = self.audio.load_wav(record.wav_path)
wav = self.audio.trim(wav)
mel_spectrogram = self.audio.melspectrogram(wav).astype(np.float32).T
file_path = os.path.join(self.out_dir, f"{record.key}.target.tfrecord")
write_preprocessed_target_data(record.id, record.key, mel_spectrogram,
file_path)
return MelStatistics(id=record.id,
key=record.key,
min=np.min(mel_spectrogram, axis=0),
max=np.max(mel_spectrogram, axis=0),
sum=np.sum(mel_spectrogram, axis=0),
length=len(mel_spectrogram),
moment2=np.sum(np.square(mel_spectrogram),
axis=0))
def _process_txt(self, record: TxtWavRecord):
with open(os.path.join(self.in_dir, record.txt_path),
mode='r',
encoding='utf8') as f:
txt = f.readline().rstrip("\n")
sequence, clean_text = text_to_sequence(txt, basic_cleaners)
source = np.array(sequence, dtype=np.int64)
file_path = os.path.join(self.out_dir,
f"{record.key}.source.tfrecord")
write_preprocessed_source_data(record.id, record.key, source,
clean_text, record.speaker_info.id,
record.speaker_info.age,
record.speaker_info.gender,
file_path)
return record.key
class LJSpeech:
def __init__(self, in_dir, out_dir, hparams):
self.in_dir = in_dir
self.out_dir = 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.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_targets(self, rdd: RDD):
return TargetRDD(
rdd.mapValues(self._process_target).persist(
StorageLevel.MEMORY_AND_DISK))
def process_sources(self, rdd: RDD):
return rdd.mapValues(self._process_source)
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 _text_to_sequence(self, text):
sequence, clean_text = text_to_sequence(text, english_cleaners)
sequence = np.array(sequence, dtype=np.int64)
return sequence, clean_text
def _process_target(self, paths: TextAndPath):
wav = self.audio.load_wav(paths.wav_path)
mel_spectrogram = self.audio.melspectrogram(wav).astype(np.float32).T
filename = f"{paths.key}.target.tfrecord"
filepath = os.path.join(self.out_dir, filename)
write_preprocessed_target_data(paths.id, paths.key, mel_spectrogram,
filepath)
return MelStatistics(id=paths.id,
key=paths.key,
min=np.min(mel_spectrogram, axis=0),
max=np.max(mel_spectrogram, axis=0),
sum=np.sum(mel_spectrogram, axis=0),
length=len(mel_spectrogram),
moment2=np.sum(np.square(mel_spectrogram),
axis=0))
def _process_source(self, paths: TextAndPath):
sequence, clean_text = self._text_to_sequence(paths.text)
filename = f"{paths.key}.source.tfrecord"
filepath = os.path.join(self.out_dir, filename)
write_preprocessed_source_data(paths.id, paths.key, sequence,
clean_text, filepath)
return paths.key
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))
