def split_train_validation_data():
traindata = np.load("traindata.npz", allow_pickle=True)['arr_0']
male_list = []
female_list = []
record_id = []
for i in range(len(traindata)):
f_name = traindata[i]['filename']
gender, id_num, _, _ = path2info(f_name)
if id_num not in record_id:
record_id.append(id_num)
if gender == 'man':
male_list.append(id_num)
else:
female_list.append(id_num)
print(len(female_list), "and ", len(male_list))
train_size = (int)(len(record_id) * 0.9)
male_ratio = len(male_list) / train_size
male_train_size = (int)(train_size * male_ratio)
print(male_train_size, ' ', train_size)
male_train_ids = np.random.choice(male_list, male_train_size, replace=False)
female_train_ids = np.random.choice(female_list, train_size-male_train_size,replace=False)
train_ids = np.concatenate([male_train_ids, female_train_ids])
train_data = []
vali_data = []
for i in range(len(traindata)):
_, id_num, _, _ = path2info(traindata[i]['filename'])
if id_num in train_ids:
train_data.append(traindata[i])
else:
vali_data.append(traindata[i])
print('num of male speaker for training:', len(male_train_ids))
print('num of female speaker for training:', len(female_train_ids))
print('num of total train data:', len(train_data))
print('num of total val data:', len(vali_data))
np.save('train_split_data.npy', train_data)
np.save('vali_split_data.npy', vali_data)
def feature_extraction_and_force_alignment(filepath, nstates, phoneHMMs):
"""
handle one .wav file
"""
samples, samplingrate = loadAudio(filepath)
wordTrans = list(path2info(filepath)[2])
phoneTrans = words2phones(wordTrans, prondict)
stateTrans = [phone + '_' + str(stateid) for phone in phoneTrans
for stateid in range(nstates[phone])]
lmfcc_result = mfcc(samples)
mspec_result = mspec(samples)
targets = []
_, viterbi_path = forcedAlignment(lmfcc_result, phoneHMMs, phoneTrans)
targets = [stateTrans[idx] for idx in viterbi_path.astype(np.int16)]
return lmfcc_result, mspec_result, targets
# E.g. # of states of ah = 3; ah -> ['ah_0', 'ah_1', 'ah_2']
stateList = list()
for ph in phoneHMMs.keys():
for i in range(nstates[ph]):
stateList.append('%s_%d' % (ph, i))
# --------------------------------------------------------------
data = list()
for root, dirs, files in walk(folder_to_extract):
for f in tqdm(files):
if not f.endswith('.wav'):
continue
# do our work
filename = os.path.join(root, f)
sample, srate = loadAudio(filename)
mspec_x = mspec(sample, samplingrate=srate)
lmfcc_x = mfcc(sample, samplingrate=srate)
wordTrans = list(path2info(filename)[2])
phoneTrans = words2phones(wordTrans, prondict)
targets = forcedAlignment(lmfcc_x, phoneHMMs, phoneTrans)
# convert the targets from str to int
idx_targets = [stateList.index(t) for t in targets]
data.append({
'filename': filename,
'lmfcc': lmfcc_x,
'mspec': mspec_x,
'targets': idx_targets
})
kwargs = {data_type: data}
np.savez(dump_file_name, **kwargs)
sumTot = 0.0
N = len(wordTest['transmat'])
for i in range(N):
sumTot += np.sum(wordTest['transmat'][i])
#print("les deux nombres suivants doivent etre egaux")
print(sumTot)
print(N)
#print(wordTest['startprob'])
#print(wordTest['transmat'])
###########
filename = 'tidigits/disc_4.1.1/tidigits/train/man/nw/z43a.wav'
samples, samplingrate = lab3_tools.loadAudio(filename)
lmfcc = lab1.mfcc(samples)
wordTrans = list(lab3_tools.path2info(filename)[2])
print(wordTrans)
#should be ['z', '4', '3']
phoneTrans = words2phones(wordTrans, prondict, addShortPause=True)
print(phoneTrans)
#should be ['sil', 'z', 'iy', 'r', 'ow', 'f', 'ao', 'r', 'th', 'r', 'iy', 'sil']
stateTrans = [
phone + '_' + str(stateid) for phone in phoneTrans
for stateid in range(nstates[phone])
]
print(stateTrans)
wordHMMs = concatAnyHMM(phoneHMMs, phoneTrans)
import numpy as np
import json
from lab3_tools import path2info
from tqdm import tqdm
if __name__ == "__main__":
# load the validation speaker list; from the result of the script:
# work/get_val_spk_json.py
with open('data/val_spk.json', 'r') as f:
data = json.load(f)
val_spk = data['valdation_spker']
# load the train data
traindata = np.load('data/traindata.npz')['traindata']
train = list()
validation = list()
for d in tqdm(traindata):
sound_fname = d['filename']
gender, speakerID, _, _ = path2info(sound_fname)
spk_gen_id = '{}-{}'.format(gender, speakerID)
if spk_gen_id in val_spk:
validation.append(d)
else:
train.append(d)
# ===================================
print("The size of training set : ", len(train))
print("The size of validation set : ", len(validation))
kwargs = {'validation': validation, 'train': train}
np.savez('data/train_val_data.npz', **kwargs)