def wavgen_magphase(gen_dir, file_id_list, cfg, logger):
# Import MagPhase and libraries:
sys.path.append(cfg.magphase_bindir)
import libutils as lu
import libaudio as la
import magphase as mp
nfiles = len(file_id_list)
for nxf in xrange(nfiles):
filename_token = file_id_list[nxf]
logger.info('Creating waveform for %4d of %4d: %s' %
(nxf + 1, nfiles, filename_token))
for pf_type in cfg.magphase_pf_type:
gen_wav_dir = os.path.join(gen_dir + '_wav_pf_' + pf_type)
lu.mkdir(gen_wav_dir)
mp.synthesis_from_acoustic_modelling(
gen_dir,
filename_token,
gen_wav_dir,
cfg.mag_dim,
cfg.real_dim,
cfg.sr,
pf_type=pf_type,
b_const_rate=cfg.magphase_const_rate)
return
def convert(file_id_list,
in_lab_dir,
in_feats_dir,
fs,
out_lab_dir,
b_prevent_zeros=False):
'''
b_prevent_zeros: True if you want to ensure that all the phonemes have one frame at least.
(not recommended, only useful when there are too many utterances crashed)
'''
# Conversion:
lu.mkdir(out_lab_dir)
v_filenames = lu.read_text_file2(file_id_list,
dtype='string',
comments='#')
crashlist_file = lu.ins_pid('crash_file_list.scp')
for filename in v_filenames:
# Display:
print('\nConverting lab file: ' + filename +
'................................')
# Current i/o files:
in_lab_file = os.path.join(in_lab_dir, filename + '.lab')
out_lab_file = os.path.join(out_lab_dir, filename + '.lab')
in_shift_file = os.path.join(in_feats_dir, filename + '.shift')
# Debug:
'''
v_shift = lu.read_binfile(in_shift_file, dim=1)
v_n_frms = mp.get_num_of_frms_per_state(v_shift, in_lab_file, fs, b_prevent_zeros=b_prevent_zeros)
la.convert_label_state_align_to_var_frame_rate(in_lab_file, v_n_frms, out_lab_file)
#'''
v_n_frms = 0
try:
v_shift = lu.read_binfile(in_shift_file, dim=1)
v_n_frms = mp.get_num_of_frms_per_state(
v_shift,
in_lab_file,
fs,
b_prevent_zeros=b_prevent_zeros,
n_states_x_phone=1)
la.convert_label_state_align_to_var_frame_rate(
in_lab_file, v_n_frms, out_lab_file)
except (KeyboardInterrupt, SystemExit):
raise
except:
print("crashlist")
with open(crashlist_file, "a") as crashlistlog:
crashlistlog.write(filename + '\n')
print('Done!')
def convert(file_id_list, in_lab_dir, in_feats_dir, fs, out_lab_dir, b_prevent_zeros=False):
'''
b_prevent_zeros: True if you want to ensure that all the phonemes have one frame at least.
(not recommended, only useful when there are too many utterances crashed)
'''
# Conversion:
lu.mkdir(out_lab_dir)
v_filenames = lu.read_text_file2(file_id_list, dtype='string', comments='#')
crashlist_file = lu.ins_pid('crash_file_list.scp')
for filename in v_filenames:
# Display:
print('\nConverting lab file: ' + filename + '................................')
# Current i/o files:
in_lab_file = os.path.join(in_lab_dir , filename + '.lab')
out_lab_file = os.path.join(out_lab_dir , filename + '.lab')
in_shift_file = os.path.join(in_feats_dir, filename + '.shift')
# Debug:
'''
v_shift = lu.read_binfile(in_shift_file, dim=1)
v_n_frms = mp.get_num_of_frms_per_state(v_shift, in_lab_file, fs, b_prevent_zeros=b_prevent_zeros)
la.convert_label_state_align_to_var_frame_rate(in_lab_file, v_n_frms, out_lab_file)
#'''
try:
v_shift = lu.read_binfile(in_shift_file, dim=1)
v_n_frms = mp.get_num_of_frms_per_state(v_shift, in_lab_file, fs, b_prevent_zeros=b_prevent_zeros)
la.convert_label_state_align_to_var_frame_rate(in_lab_file, v_n_frms, out_lab_file)
except (KeyboardInterrupt, SystemExit):
raise
except:
with open(crashlist_file, "a") as crashlistlog:
crashlistlog.write(filename + '\n')
print('Done!')
def wavgen_magphase(gen_dir, file_id_list, cfg, logger):
# Import MagPhase and libraries:
sys.path.append(cfg.magphase_bindir)
import libutils as lu
import libaudio as la
import magphase as mp
nfiles = len(file_id_list)
for nxf in xrange(nfiles):
filename_token = file_id_list[nxf]
logger.info('Creating waveform for %4d of %4d: %s' % (nxf+1, nfiles, filename_token))
for pf_type in cfg.magphase_pf_type:
gen_wav_dir = os.path.join(gen_dir + '_wav_pf_' + pf_type)
lu.mkdir(gen_wav_dir)
mp.synthesis_from_acoustic_modelling(gen_dir, filename_token, gen_wav_dir, cfg.mag_dim, cfg.real_dim,
cfg.sr, pf_type=pf_type, b_const_rate=cfg.magphase_const_rate)
return
import magphase as mp
if __name__ == '__main__':
# CONSTANTS: So far, the vocoder has been tested only with the following constants:===
fs = 48000
# INPUT:==============================================================================
files_scp = '../data/file_id.scp' # List of file names (tokens). Format used by Merlin.
in_lab_st_dir = '../data/labs' # Original state aligned label files directory (in the format used by Merlin).
in_shift_dir = '../data/params' # Directory containing .shift files (You need to run feature extraction before running this script.)
out_lab_st_dir = '../data/labs_var_rate' # Directory that will contain the converted "variable frame rate" state aligned label files.
b_prevent_zeros = False # True if you want to ensure that all the phonemes have one frame at least. (not recommended, only usful when there are too many utterances crashed)
# PROCESSING:=========================================================================
lu.mkdir(out_lab_st_dir)
v_fileTokns = lu.read_text_file2(files_scp, dtype='string', comments='#')
n_files = len(v_fileTokns)
crashlist_file = lu.ins_pid('crash_file_list.scp')
for ftkn in v_fileTokns:
# Display:
print('\nAnalysing file: ' + ftkn + '................................')
# Input files:
in_lab_st_file = in_lab_st_dir + '/' + ftkn + '.lab'
out_lab_st_file = out_lab_st_dir + '/' + ftkn + '.lab'
in_shift_file = in_shift_dir + '/' + ftkn + '.shift'
try:
if __name__ == '__main__':
# CONSTANTS: So far, the vocoder has been tested only with the following constants:===
fs = 48000
# INPUT:==============================================================================
files_scp = '../data_48k/file_id.scp' # List of file names (tokens). Format used by Merlin.
in_lab_st_dir = '../data_48k/labs' # Original state aligned label files directory (in the format used by Merlin).
in_shift_dir = '../data_48k/params' # Directory containing .shift files (You need to run feature extraction before running this script.)
out_lab_st_dir = '../data_48k/labs_var_rate' # Directory that will contain the converted "variable frame rate" state aligned label files.
b_prevent_zeros = False # True if you want to ensure that all the phonemes have one frame at least. (not recommended, only usful when there are too many utterances crashed)
# PROCESSING:=========================================================================
lu.mkdir(out_lab_st_dir)
v_fileTokns = lu.read_text_file2(files_scp, dtype='string', comments='#')
n_files = len(v_fileTokns)
crashlist_file = lu.ins_pid('crash_file_list.scp')
for ftkn in v_fileTokns:
# Display:
print('\nAnalysing file: ' + ftkn + '................................')
# Input files:
in_lab_st_file = in_lab_st_dir + '/' + ftkn + '.lab'
out_lab_st_file = out_lab_st_dir + '/' + ftkn + '.lab'
in_shift_file = in_shift_dir + '/' + ftkn + '.shift'
try:
lu.write_binfile(v_lf0, out_feats_dir + '/' + file_name_token + '.lf0')
# Saving auxiliary feature shift (hop length). It is useful for posterior modifications of labels in Merlin.
lu.write_binfile(v_shift, out_feats_dir + '/' + file_name_token + '.shift')
return
if __name__ == '__main__':
# CONSTANTS: So far, the vocoder has been tested only with the following constants:===
fft_len = 4096
fs = 48000
# INPUT:==============================================================================
files_scp = '../data/file_id.scp' # List of file names (tokens). Format used by Merlin.
in_wav_dir = '../data/wavs_nat' # Directory with the wavfiles to extract the features from.
out_feats_dir = '../data/params' # Output directory that will contain the extracted features.
mvf = 4500 # Maximum voiced frequency (Hz)
# FILES SETUP:========================================================================
lu.mkdir(out_feats_dir)
l_file_tokns = lu.read_text_file2(files_scp, dtype='string',
comments='#').tolist()
# MULTIPROCESSING EXTRACTION:==========================================================
lu.run_multithreaded(feat_extraction, in_wav_dir, l_file_tokns,
out_feats_dir, fft_len, mvf)
print('Done!')
copy2(
join(this_dir, 'conf_base', 'logging_config.conf'),
join(exper_path, 'acoustic_model', 'conf', 'logging_config.conf'))
# Read file list:
file_id_list = pars_acous_train['Paths']['file_id_list']
l_file_tokns = lu.read_text_file2(file_id_list,
dtype='string',
comments='#').tolist()
acoustic_feats_path = pars_acous_train['Paths']['in_acous_feats_dir']
# Acoustic Feature Extraction:-------------------------------------------------------------
if b_feat_extr:
# Extract features:
lu.mkdir(acoustic_feats_path)
if b_feat_ext_multiproc:
lu.run_multithreaded(
feat_extraction,
join(exper_path, 'acoustic_model', 'data', 'wav'),
l_file_tokns, acoustic_feats_path, d_mp_opts)
else:
for file_name_token in l_file_tokns:
feat_extraction(
join(exper_path, 'acoustic_model', 'data', 'wav'),
file_name_token, acoustic_feats_path, d_mp_opts)
# Labels Conversion to Variable Frame Rate:------------------------------------------------
if b_conv_labs_rate and not d_mp_opts[
'b_const_rate']: # NOTE: The script ./script/label_st_align_to_var_rate.py can be also called from comand line directly.
lp.xlabel('Time (frames)')
lp.ylabel('F0')
lp.grid()
return
if __name__ == '__main__':
# INPUT:==============================================================================
wav_file_orig = 'data_48k/wavs_nat/hvd_593.wav' # Original natural waveform. You can choose any of the provided ones in the /wavs_nat directory.
out_dir = 'data_48k/wavs_syn' # Where the synthesised waveform will be stored
b_plots = True # True if you want to plot the extracted parameters.
# PROCESS:============================================================================
lu.mkdir(out_dir)
# ANALYSIS:
print("Analysing.....................................................")
m_mag, m_real, m_imag, v_f0, fs, v_shift = mp.analysis_lossless(
wav_file_orig)
# MODIFICATIONS:
# You can modify the parameters here if wanted.
# SYNTHESIS:
print("Synthesising.................................................")
v_syn_sig = mp.synthesis_from_lossless(m_mag, m_real, m_imag, v_f0, fs)
# SAVE WAV FILE:
print("Saving wav file..............................................")
# File setup:
wav_file = os.path.join(in_wav_dir, file_name_token + '.wav')
mp.analysis_compressed(wav_file, out_dir=out_feats_dir)
return
if __name__ == '__main__':
# INPUT:==============================================================================
files_scp = '../data_48k/file_id.scp' # List of file names (tokens). Format used by Merlin.
in_wav_dir = '../data_48k/wavs_nat' # Directory with the wavfiles to extract the features from.
out_feats_dir = '../data_48k/params' # Output directory that will contain the extracted features.
# FILES SETUP:========================================================================
lu.mkdir(out_feats_dir)
l_file_tokns = lu.read_text_file2(files_scp, dtype='string', comments='#').tolist()
# MULTIPROCESSING EXTRACTION:==========================================================
lu.run_multithreaded(feat_extraction, in_wav_dir, l_file_tokns, out_feats_dir)
# For debug (Don't remove):
#for file_name_token in l_file_tokns:
# feat_extraction(in_wav_dir, file_name_token, out_feats_dir)
print('Done!')
os.path.join(exper_path, 'conf/config_base.conf'))
# Save backup of this file and used magphase code:
shutil.copytree(os.path.dirname(mp.__file__),
os.path.join(exper_path, 'backup_magphase_code'))
shutil.copy2(__file__, os.path.join(exper_path, 'conf'))
# Read file list:
l_file_tokns = lu.read_text_file2(os.path.join(exper_path, file_id_list),
dtype='string',
comments='#').tolist()
if b_feat_extr:
# Extract features:
acoustic_feats_path = os.path.join(exper_path, acoustic_feats_dir)
lu.mkdir(acoustic_feats_path)
if b_feat_ext_multiproc:
lu.run_multithreaded(feat_extraction, in_wav_dir, l_file_tokns,
acoustic_feats_path, d_mp_opts)
else:
for file_name_token in l_file_tokns:
feat_extraction(in_wav_dir, file_name_token,
acoustic_feats_path, d_mp_opts)
if b_config_merlin or b_wavgen:
# Edit Merlin's config file:
parser = configparser.ConfigParser()
parser.optionxform = str
parser.read([os.path.join(exper_path, 'conf/config_base.conf')])
# INPUT:==============================================================================
files_scp = '../demos/data_48k/file_id_predict.scp' # List of file names (tokens). Format used by Merlin.
in_feats_dir = '../demos/data_48k/params_predicted' # Input directory that contains the predicted features.
out_syn_dir = '../demos/data_48k/wavs_syn_from_predicted' # Where the synthesised waveform will be stored.
mag_dim = 60 # Number of Mel-scaled frequency bins.
phase_dim = 45 # Number of Mel-scaled frequency bins kept for phase features (real and imag). It must be <= mag_dim
pf_type = 'magphase' # "magphase": MagPhase's own postfilter (in development)
# "merlin": Merlin's style postfilter.
# "no": No postfilter.
b_multiproc = False # If True, it synthesises using all the available cores in parallel. If False, it just uses one core (slower).
# FILES SETUP:========================================================================
lu.mkdir(out_syn_dir)
l_file_tokns = lu.read_text_file2(files_scp, dtype='string',
comments='#').tolist()
# PROCESSING:=========================================================================
if b_multiproc:
lu.run_multithreaded(synthesis, in_feats_dir, l_file_tokns,
out_syn_dir, mag_dim, phase_dim, fs, pf_type)
else:
for file_tokn in l_file_tokns:
synthesis(in_feats_dir, file_tokn, out_syn_dir, mag_dim, phase_dim,
fs, pf_type)
print('Done!')
if __name__ == '__main__':
# Parsing input arg:
config_file = sys.argv[1]
# Constants:
b_prevent_zeros = False # True if you want to ensure that all the phonemes have one frame at least.
# (not recommended, only usful when there are too many utterances crashed)
# Parsing config file:
file_id_list, in_lab_dir, in_feats_dir, fs, out_lab_dir = parse_config_file(config_file)
# Conversion:
lu.mkdir(out_lab_dir)
v_filenames = lu.read_text_file2(file_id_list, dtype='string', comments='#')
n_files = len(v_filenames)
crashlist_file = lu.ins_pid('crash_file_list.scp')
for filename in v_filenames:
# Display:
print('\nConverting lab file: ' + filename + '................................')
# Current i/o files:
in_lab_file = path.join(in_lab_dir , filename + '.lab')
out_lab_file = path.join(out_lab_dir , filename + '.lab')
# Debug:
lp.xlabel('Time (frames)')
lp.ylabel('F0')
lp.grid()
return
if __name__ == '__main__':
# INPUT:==============================================================================
wav_file_orig = 'data_48k/wavs_nat/hvd_593.wav' # Original natural waveform. You can choose any of the provided ones in the /wavs_nat directory.
out_dir = 'data_48k/wavs_syn' # Where the synthesised waveform will be stored
b_plots = True # True if you want to plot the extracted parameters.
# PROCESS:============================================================================
lu.mkdir(out_dir)
# ANALYSIS:
print("Analysing.....................................................")
m_mag, m_real, m_imag, v_f0, fs, v_shift = mp.analysis_lossless(wav_file_orig)
# MODIFICATIONS:
# You can modify the parameters here if wanted.
# SYNTHESIS:
print("Synthesising.................................................")
v_syn_sig = mp.synthesis_from_lossless(m_mag, m_real, m_imag, v_f0, fs)
# SAVE WAV FILE:
print("Saving wav file..............................................")
wav_file_syn = out_dir + '/' + lu.get_filename(wav_file_orig) + '_copy_syn_lossless.wav'
fs = 48000
# INPUT:==============================================================================
files_scp = '../data_48k/file_id.scp' # List of file names (tokens). Format used by Merlin.
in_feats_dir = '../data_48k/params' # Input directory that contains the predicted features.
out_syn_dir = '../data_48k/wavs_syn_merlin' # Where the synthesised waveform will be stored.
nbins_mel = 60 # Number of Mel-scaled frequency bins.
nbins_phase = 45 # Number of Mel-scaled frequency bins kept for phase features (real and imag). It must be <= nbins_mel
b_postfilter = True # If True, the MagPhase vocoder post-filter is applied. Note: If you want to use the one included in Merlin, disable this one.
b_parallel = False # If True, it synthesises using all the available cores in parallel. If False, it just uses one core (slower).
# FILES SETUP:========================================================================
lu.mkdir(out_syn_dir)
l_file_tokns = lu.read_text_file2(files_scp, dtype='string', comments='#').tolist()
# PROCESSING:=========================================================================
if b_parallel:
lu.run_multithreaded(synthesis, in_feats_dir, l_file_tokns, out_syn_dir, nbins_mel, nbins_phase, fs, b_postfilter)
else:
for file_tokn in l_file_tokns:
synthesis(in_feats_dir, file_tokn, out_syn_dir, nbins_mel, nbins_phase, fs, b_postfilter)
print('Done!')
save_config(pars_dur_train, join(dur_model_conf_path , 'dur_train.conf'))
save_config(pars_dur_synth, join(dur_model_conf_path , 'dur_synth.conf'))
save_config(pars_acous_train, join(acous_model_conf_path, 'acous_train.conf'))
save_config(pars_acous_synth, join(acous_model_conf_path, 'acous_synth.conf'))
copy2(join(this_dir, 'conf_base', 'logging_config.conf'), join(exper_path, 'acoustic_model', 'conf', 'logging_config.conf'))
# Read file list:
file_id_list = pars_acous_train['Paths']['file_id_list']
l_file_tokns = lu.read_text_file2(file_id_list, dtype='string', comments='#').tolist()
acoustic_feats_path = pars_acous_train['Paths']['in_acous_feats_dir']
# Acoustic Feature Extraction:-------------------------------------------------------------
if b_feat_extr:
# Extract features:
lu.mkdir(acoustic_feats_path)
if b_feat_ext_multiproc:
lu.run_multithreaded(feat_extraction, join(exper_path, 'acoustic_model', 'data', 'wav'), l_file_tokns, acoustic_feats_path, d_mp_opts)
else:
for file_name_token in l_file_tokns:
feat_extraction(join(exper_path, 'acoustic_model', 'data', 'wav'), file_name_token, acoustic_feats_path, d_mp_opts)
# Labels Conversion to Variable Frame Rate:------------------------------------------------
if b_conv_labs_rate and not d_mp_opts['b_const_rate']: # NOTE: The script ./script/label_st_align_to_var_rate.py can be also called from comand line directly.
label_state_align = join(exper_path, 'acoustic_model', 'data', 'label_state_align')
label_state_align_var_rate = pars_acous_train['Labels']['label_align']
fs = int(pars_acous_train['Waveform']['samplerate'])
ltvr.convert(file_id_list,label_state_align, acoustic_feats_path, fs, label_state_align_var_rate)
# Run duration training:-------------------------------------------------------------------
