def analysis_for_valid_batch(self,
features,
output_features,
out_dir,
sample_rate=16000,
**kwargs):
kwargs['sample_rate'] = sample_rate
super(LSTMAcousticModel,
self).analysis_for_valid_batch(features, output_features,
out_dir, **kwargs)
# Synthesise outputs using WORLD.
synth_dir = os.path.join(out_dir, 'synth')
os.makedirs(synth_dir, exist_ok=True)
lf0, vuv, mcep, bap = utils.detach_batched_seqs(
output_features['lf0'],
output_features['vuv'],
output_features['mcep'],
output_features['bap'],
seq_len=features['n_frames'])
for _lf0, _vuv, _mcep, _bap, _file_id in zip(lf0, vuv, mcep, bap,
features['name']):
_vuv = _vuv > 0.5
_f0 = np.exp(_lf0)
_f0 = savgol_filter(_f0, 7, 1)
wav_path = os.path.join(synth_dir, f'{_file_id}.wav')
wav = world_with_reaper_f0.synthesis(_f0, _vuv, _mcep, _bap,
sample_rate)
file_io.save_wav(wav, wav_path, sample_rate=sample_rate)
def analysis_for_valid_batch(self, features, output_features, out_dir, sample_rate=16000, **kwargs):
kwargs['sample_rate'] = sample_rate
super(F0Model, self).analysis_for_valid_batch(features, output_features, out_dir, **kwargs)
# Synthesise outputs using WORLD.
synth_dir = os.path.join(out_dir, 'synth')
os.makedirs(synth_dir, exist_ok=True)
lf0 = output_features['lf0'].cpu().detach().numpy()
vuv = features['vuv'].cpu().detach().numpy()
sp = features['sp'].cpu().detach().numpy()
ap = features['ap'].cpu().detach().numpy()
n_frames = features['n_frames'].cpu().detach().numpy()
for i, (n_frame, name) in enumerate(zip(n_frames, features['name'])):
f0_i = np.exp(lf0[i, :n_frame, 0])
f0_i = savgol_filter(f0_i, 7, 1)
f0_i = f0_i * vuv[i, :n_frame, 0]
f0_i = f0_i.astype(np.float64)
sp_i = sp[i, :n_frame].astype(np.float64)
ap_i = ap[i, :n_frame].astype(np.float64)
wav_path = os.path.join(synth_dir, '{}.wav'.format(name))
wav = pyworld.synthesize(f0_i, sp_i, ap_i, sample_rate)
file_io.save_wav(wav, wav_path, sample_rate=sample_rate)
def process(wav_dir, id_list, out_dir, calculate_normalisation,
normalisation_of_deltas):
"""Processes wav files in id_list, saves the log-F0 and MVN parameters to files.
Args:
wav_dir (str): Directory containing the wav files.
id_list (str): List of file basenames to process.
out_dir (str): Directory to save the output to.
calculate_normalisation (bool): Whether to automatically calculate MVN parameters after extracting F0.
normalisation_of_deltas (bool): Also calculate the MVN parameters for the delta and delta delta features.
"""
file_ids = get_file_ids(id_list=id_list)
make_dirs(os.path.join(out_dir, 'lf0'), file_ids)
make_dirs(os.path.join(out_dir, 'vuv'), file_ids)
make_dirs(os.path.join(out_dir, 'mcep'), file_ids)
make_dirs(os.path.join(out_dir, 'bap'), file_ids)
make_dirs(os.path.join(out_dir, 'wav_synth'), file_ids)
for file_id in file_ids:
wav_path = os.path.join(wav_dir, f'{file_id}.wav')
wav, sample_rate = file_io.load_wav(wav_path)
f0, vuv, mcep, bap = analysis(wav, sample_rate)
lf0 = np.log(f0)
wav_synth = synthesis(f0, vuv, mcep, bap, sample_rate)
file_io.save_bin(lf0, os.path.join(out_dir, 'lf0', f'{file_id}.npy'))
file_io.save_bin(vuv, os.path.join(out_dir, 'vuv', f'{file_id}.npy'))
file_io.save_bin(mcep, os.path.join(out_dir, 'mcep', f'{file_id}.npy'))
file_io.save_bin(bap, os.path.join(out_dir, 'bap', f'{file_id}.npy'))
file_io.save_wav(wav_synth,
os.path.join(out_dir, 'wav_synth', f'{file_id}.wav'),
sample_rate)
if calculate_normalisation:
process_mvn(out_dir,
'lf0',
id_list=id_list,
deltas=normalisation_of_deltas,
out_dir=out_dir)
process_mvn(out_dir,
'mcep',
id_list=id_list,
deltas=normalisation_of_deltas,
out_dir=out_dir)
process_mvn(out_dir,
'bap',
id_list=id_list,
deltas=normalisation_of_deltas,
out_dir=out_dir)
def save_file(self, data, base_name, data_dir):
r"""Saves the feature as a wavfile using scipy.io.wavfile.write
Parameters
----------
data : int or float or bool or `np.ndarray`, shape (seq_len, feat_dim)
Data loaded from the file specified.
base_name : str
The name (without extensions) of the file to be loaded.
data_dir : str
The directory containing all feature types for this dataset.
"""
file_path = self.file_path(base_name, data_dir)
file_io.save_wav(data, file_path, self.sample_rate)
def batch_synth(features, output_features, names, out_dir, sample_rate=16000):
synth_dir = os.path.join(out_dir, 'synth')
os.makedirs(synth_dir, exist_ok=True)
lf0, vuv, sp, ap = utils.detach_batched_seqs(output_features['lf0'],
features['vuv'],
features['sp'],
features['ap'],
seq_len=features['n_frames'])
for i, name in enumerate(names):
f0_i = np.exp(lf0[i])
f0_i = savgol_filter(f0_i, 7, 1)
wav_path = os.path.join(synth_dir, '{}.wav'.format(name))
wav = world.synthesis(f0_i,
vuv[i],
sp[i],
ap[i],
sample_rate=sample_rate)
file_io.save_wav(wav, wav_path, sample_rate=sample_rate)
def batch_synth(lf0,
vuv,
mcep,
bap,
seq_len=None,
names=None,
out_dir=None,
sample_rate=16000):
if out_dir is not None:
if names is None:
raise ValueError(
'If `out_dir` is given, then `names` of individual sentences must also be given'
)
synth_dir = os.path.join(out_dir, 'synth')
make_dirs(synth_dir, names)
lf0, vuv, mcep, bap = utils.detach_batched_seqs(lf0,
vuv,
mcep,
bap,
seq_len=seq_len)
wavs = []
for i, name in enumerate(names):
f0_i = np.exp(lf0[i])
f0_i = savgol_filter(f0_i, 7, 1) if len(f0_i) >= 7 else f0_i
wav = world.synthesis(f0_i,
vuv[i],
mcep[i],
bap[i],
sample_rate=sample_rate)
wavs.append(wav)
if out_dir is not None:
wav_path = os.path.join(synth_dir, f'{names[i]}.wav')
file_io.save_wav(wav, wav_path, sample_rate=sample_rate)
return wavs