def Preprocessing_general_speaker(self):
"""
Go through the sentences one by one.
- reads ema data and turn it to a (K,18) array where arti are in a precise order, interploate missing values,
smooth the trajectories, remove silences at the beginning and the end, undersample to have 1 position per
frame mfcc, add it to the list of EMA traj for this speaker
- reads the wav file, calculate the associated acoustic features (mfcc+delta+ deltadelta+contextframes) ,
add it to the list of the MFCC FEATURES for this speaker.
Then calculate the normvalues based on the list of ema/mfcc data for this speaker
Finally : normalization and last smoothing of the trajectories.
Final data are in Preprocessed_data/speaker/ema_final.npy and mfcc.npy
"""
self.create_missing_dir()
N = len(self.EMA_files)
if self.N_max != 0:
N = self.N_max
for i in range(N):
ema = self.read_ema_file(i)
ema_VT = self.add_vocal_tract(ema)
ema_VT_smooth = self.smooth_data(ema_VT)
path_wav = os.path.join(self.path_wav_files,
self.EMA_files[i] + '.wav')
wav, sr = librosa.load(path_wav, sr=self.sampling_rate_wav)
wav = 0.5 * wav / np.max(wav)
mfcc = self.from_wav_to_mfcc(wav)
ema_VT_smooth, mfcc = self.remove_silences(i, ema_VT_smooth, mfcc)
ema_VT_smooth, mfcc = self.synchro_ema_mfcc(ema_VT_smooth, mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema", self.EMA_files[i]), ema_VT)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"mfcc", self.EMA_files[i]), mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema_final", self.EMA_files[i]), ema_VT_smooth)
self.list_EMA_traj.append(ema_VT_smooth)
self.list_MFCC_frames.append(mfcc)
self.calculate_norm_values()
for i in range(N):
ema_VT_smooth = np.load(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema_final", self.EMA_files[i] + ".npy"))
mfcc = np.load(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"mfcc", self.EMA_files[i] + ".npy"))
ema_VT_smooth_norma, mfcc = self.normalize_sentence(
i, ema_VT_smooth, mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"mfcc", self.EMA_files[i]), mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema_final", self.EMA_files[i]),
ema_VT_smooth_norma)
# split_sentences(speaker)
get_fileset_names(self.speaker)
def Preprocessing_general_speaker(self):
self.create_missing_dir()
N = len(self.EMA_files)
if self.N_max != 0:
N = self.N_max
for i in range(N):
ema = self.read_ema_file(i)
ema_VT = self.add_vocal_tract(ema)
ema_VT_smooth = self.smooth_data(ema_VT)
path_wav = os.path.join(self.path_wav_files,
self.EMA_files[i] + '.wav')
wav, sr = librosa.load(path_wav, sr=self.sampling_rate_wav)
wav = 0.5 * wav / np.max(wav)
mfcc = self.from_wav_to_mfcc(wav)
ema_VT_smooth, mfcc = self.remove_silences(self.EMA_files[i],
ema_VT_smooth, mfcc)
n_frames_wanted = mfcc.shape[0]
ema_VT_smooth = scipy.signal.resample(ema, num=n_frames_wanted)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema", self.EMA_files[i]), ema_VT)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"mfcc", self.EMA_files[i]), mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema_final", self.EMA_files[i]), ema_VT_smooth)
self.list_EMA_traj.append(ema_VT_smooth)
self.list_MFCC_frames.append(mfcc)
self.calculate_norm_values()
for i in range(N):
ema_VT_smooth = np.load(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema_final", self.EMA_files[i] + ".npy"))
mfcc = np.load(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"mfcc", self.EMA_files[i] + ".npy"))
ema_VT_smooth_norma, mfcc = self.normalize_sentence(
i, ema_VT_smooth, mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"mfcc", self.EMA_files[i]), mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema_final", self.EMA_files[i]),
ema_VT_smooth_norma)
# split_sentences(speaker)
get_fileset_names(self.speaker)
def Preprocessing_general_speaker(self):
"""
Go through the sentences one by one.
- reads ema data and turn it to a (K,18) array where arti are in a precise order, interploate missing values,
smooth the trajectories, remove silences at the beginning and the end, undersample to have 1 position per
frame mfcc, add it to the list of EMA traj for this speaker
- reads the wav file, calculate the associated acoustic features (mfcc+delta+ deltadelta+contextframes) ,
add it to the list of the MFCC FEATURES for this speaker.
Then calculate the normvalues based on the list of ema/mfcc data for this speaker
Finally : normalization and last smoothing of the trajectories.
Final data are in Preprocessed_data/speaker/ema_final.npy and mfcc.npy
"""
self.create_missing_dir()
EMA_files = sorted([
name[:-4]
for name in os.listdir(os.path.join(self.path_files_brutes, "mat"))
if name.endswith(".mat")
])
N = len(EMA_files)
if self.N_max != 0:
N = min(
int(self.N_max / 3), N
) # majoration:if we want to preprocess N_max sentences, about N_max/6 files
self.get_data_per_sentence(
) # one file contains several sentences, this create one file per sentence
self.EMA_files_2 = sorted([
name[:-4] for name in os.listdir(
os.path.join(self.path_files_brutes, "wav_cut"))
if name.endswith(".wav")
])
N_2 = len(self.EMA_files_2)
if self.N_max != 0:
N_2 = min(self.N_max, N_2)
for i in range(N_2):
ema = self.read_ema_file(i)
ema_VT = self.add_vocal_tract(ema)
ema_VT_smooth = self.smooth_data(
ema_VT) # smooth for better calculation of norm values
mfcc = self.from_wav_to_mfcc(i)
ema_VT_smooth, mfcc = self.remove_silences(i, ema_VT_smooth, mfcc)
ema_VT_smooth, mfcc = self.synchro_ema_mfcc(ema_VT_smooth, mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema", self.EMA_files_2[i]), ema_VT)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"mfcc", self.EMA_files_2[i]), mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema_final", self.EMA_files_2[i]), ema_VT_smooth)
self.list_EMA_traj.append(ema_VT_smooth)
self.list_MFCC_frames.append(mfcc)
self.calculate_norm_values()
for i in range(N_2):
ema_VT_smooth = np.load(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema_final", self.EMA_files_2[i] + ".npy"))
mfcc = np.load(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"mfcc", self.EMA_files_2[i] + ".npy"))
ema_VT_smooth_norma, mfcc = self.normalize_sentence(
i, ema_VT_smooth, mfcc)
new_sr = 1 / self.hop_time
ema_VT_smooth_norma = self.smooth_data(ema_VT_smooth_norma, new_sr)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"mfcc", self.EMA_files_2[i]), mfcc)
np.save(
os.path.join(root_path, "Preprocessed_data", self.speaker,
"ema_final", self.EMA_files_2[i]),
ema_VT_smooth_norma)
# split_sentences(self.speaker)
get_fileset_names(self.speaker)