def sp_to_mcep(m_sp, n_coeffs=60, alpha=0.77, in_type=3, fft_len=0):
#Pre:
temp_sp = lu.ins_pid('temp.sp')
temp_mgc = lu.ins_pid('temp.mgc')
# Writing input data:
lu.write_binfile(m_sp, temp_sp)
if fft_len is 0: # case fft automatic
fft_len = 2 * (np.size(m_sp, 1) - 1)
# MCEP:
curr_cmd = _sptk_mcep_bin + " -a %1.2f -m %d -l %d -e 1.0E-8 -j 0 -f 0.0 -q %d %s > %s" % (
alpha, n_coeffs - 1, fft_len, in_type, temp_sp, temp_mgc)
call(curr_cmd, shell=True)
# Read MGC File:
m_mgc = lu.read_binfile(temp_mgc, n_coeffs)
# Deleting temp files:
os.remove(temp_sp)
os.remove(temp_mgc)
#$sptk/mcep -a $alpha -m $mcsize -l $nFFT -e 1.0E-8 -j 0 -f 0.0 -q 3 $sp_dir/$sentence.sp > $mgc_dir/$sentence.mgc
return m_mgc
def sp_to_mcep(m_sp, n_coeffs=60, alpha=0.77, in_type=3, fft_len=0):
#Pre:
temp_sp = lu.ins_pid('temp.sp')
temp_mgc = lu.ins_pid('temp.mgc')
# Writing input data:
lu.write_binfile(m_sp, temp_sp)
if fft_len is 0: # case fft automatic
fft_len = 2*(np.size(m_sp,1) - 1)
# MCEP:
curr_cmd = _sptk_mcep_bin + " -a %1.2f -m %d -l %d -e 1.0E-8 -j 0 -f 0.0 -q %d %s > %s" % (alpha, n_coeffs-1, fft_len, in_type, temp_sp, temp_mgc)
call(curr_cmd, shell=True)
# Read MGC File:
m_mgc = lu.read_binfile(temp_mgc , n_coeffs)
# Deleting temp files:
os.remove(temp_sp)
os.remove(temp_mgc)
#$sptk/mcep -a $alpha -m $mcsize -l $nFFT -e 1.0E-8 -j 0 -f 0.0 -q 3 $sp_dir/$sentence.sp > $mgc_dir/$sentence.mgc
return m_mgc
def analysis(wav_file, fft_len):
est_file = lu.ins_pid('temp.est')
la.reaper(wav_file_orig, est_file)
m_mag, m_real, m_imag, v_shift, v_voi, m_frm, fs = mp.analysis_with_del_comp__ph_enc__f0_norm__from_files_raw(wav_file, est_file, fft_len)
v_f0 = mp.shift_to_f0(v_shift, v_voi, fs, out='f0', b_smooth=True)
os.remove(est_file)
return m_mag, m_real, m_imag, v_f0
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 analysis(wav_file, fft_len, mvf, nbins_mel=60, nbins_phase=45):
est_file = lu.ins_pid('temp.est')
la.reaper(wav_file, est_file)
m_mag_mel_log, m_real_mel, m_imag_mel, v_shift, v_lf0, fs = mp.analysis_with_del_comp__ph_enc__f0_norm__from_files2(
wav_file,
est_file,
fft_len,
mvf,
f0_type='lf0',
mag_mel_nbins=nbins_mel,
cmplx_ph_mel_nbins=nbins_phase)
os.remove(est_file)
return m_mag_mel_log, m_real_mel, m_imag_mel, v_lf0
def get_pitch_marks(v_sig, fs):
temp_wav = lu.ins_pid('temp.wav')
temp_pm = lu.ins_pid('temp.pm')
sf.write(temp_wav, v_sig, fs)
reaper(temp_wav, temp_pm)
v_pm = np.loadtxt(temp_pm, skiprows=7)
v_pm = v_pm[:, 0]
# Protection against REAPER bugs 1:
vb_correct = np.hstack((True, np.diff(v_pm) > 0))
v_pm = v_pm[vb_correct]
# Protection against REAPER bugs 2 (maybe I need a better protection):
if (v_pm[-1] * fs) >= (np.size(v_sig) - 1):
v_pm = v_pm[:-1]
# Removing temp files:
os.remove(temp_wav)
os.remove(temp_pm)
return v_pm
def get_pitch_marks(v_sig, fs):
temp_wav = lu.ins_pid('temp.wav')
temp_pm = lu.ins_pid('temp.pm')
sf.write(temp_wav, v_sig, fs)
reaper(temp_wav, temp_pm)
v_pm = np.loadtxt(temp_pm, skiprows=7)
v_pm = v_pm[:,0]
# Protection against REAPER bugs 1:
vb_correct = np.hstack(( True, np.diff(v_pm) > 0))
v_pm = v_pm[vb_correct]
# Protection against REAPER bugs 2 (maybe I need a better protection):
if (v_pm[-1] * fs) >= (np.size(v_sig)-1):
v_pm = v_pm[:-1]
# Removing temp files:
os.remove(temp_wav)
os.remove(temp_pm)
return v_pm
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!')
# 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:
v_shift = lu.read_binfile(in_shift_file, dim=1)
v_n_frms = mp.get_num_of_frms_per_state(
v_shift,
in_lab_st_file,
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:
v_shift = lu.read_binfile(in_shift_file, dim=1)
v_n_frms = mp.get_num_of_frms_per_state(v_shift, in_lab_st_file, fs, b_prevent_zeros=b_prevent_zeros, n_states_x_phone=5, nfrms_tolerance=6)
# Extraction:
