def single_crm(idx_str_list, mix_path, data_path):
F_mix = np.load(mix_path)
mix_name = 'mix'
mid_name = ''
dataset_line = ''
for idx in idx_str_list:
mid_name += '-%s' % idx
mix_name += '-%s' % idx
mix_name += '.npy'
dataset_line += mix_name
for idx in idx_str_list:
single_name = 'single-%s.npy' % idx
path = '%s/single/%s' % (data_path, single_name)
F_single = np.load(path)
cRM = utils.fast_cRM(F_single, F_mix)
last_name = '-%s' % idx
cRM_name = 'crm' + mid_name + last_name + '.npy'
store_path = '%s/crm/%s' % (data_path, cRM_name)
np.save(store_path, cRM)
with open('%s/crm_log.txt' % data_path, 'a') as f:
f.write(cRM_name)
f.write('\n')
dataset_line += (' ' + cRM_name)
with open('%s/dataset.txt' % data_path, 'a') as f:
f.write(dataset_line)
f.write('\n')
def single_crm(idx_str_list, mix_path, database_repo):
F_mix = np.load(mix_path)
mid_name = ""
mix_name = "mix"
dataset_line = ""
for idx in idx_str_list:
mid_name += "-%s" % idx
mix_name += "-%s" % idx
mix_name += '.npy'
dataset_line += mix_name
for idx in idx_str_list:
single_name = 'single-%s.npy' % idx
path = '%s/single/%s' % (database_repo, single_name)
F_single = np.load(path)
cRM = utils.fast_cRM(F_single, F_mix)
last_name = '-%s' % idx
cRM_name = 'crm' + mid_name + last_name + '.npy'
# save crm to npy
store_path = '%s/crm/%s' % (database_repo, cRM_name)
np.save(store_path, cRM)
# save crm information to log
with open('%s/crm_log.txt' % database_repo, 'a') as f:
f.write(cRM_name)
f.write('\n')
dataset_line += (" " + cRM_name)
# write in database log
with open('%s/dataset.txt' % database_repo, 'a') as f:
f.write(dataset_line)
f.write('\n')
def generate_mix_sample(audio_path_list,num_speaker,fix_sr=16000,verbose=0):
'''
generate mix sample from audios in the list
:param audio_path_list: list contains path of the wav audio file
:param num_speaker: specify the task for speech separation
:param fix_sr: fix sample rate
'''
# initiate variables
# shape of F_mix = (298,257,2)
# shpae of crm = (298,257,2)
data_list = []
F_list = [] # STFT list for each sample
cRM_list = []
mix_name = "mix"
crm_name = "crm"
post_name = ""
# import data
for i in range(num_speaker):
idx,path =audio_path_list[i]
post_name += "-%05d"%idx
data, _ = librosa.load(path,sr=fix_sr)
data_list.append(data)
# create mix audio according to mix rate
mix_rate = 1.0 / float(num_speaker)
mix = np.zeros(shape=data_list[0].shape)
for data in data_list:
mix += data*mix_rate
# transfrom data via STFT and several preprocessing function
for i in range(num_speaker):
F = utils.fast_stft(data_list[i],power=False)
F_list.append(F)
F_mix = utils.fast_stft(mix,power=False)
# create cRM for each speaker and fill into y_sample
for i in range(num_speaker):
cRM_list.append(utils.fast_cRM(F_list[i],F_mix))
# return values
if verbose == 1:
print('shape of X: ',F_mix.shape)
for i in range(len(cRM_list)):
print('shape of cRM%s :'%i,cRM_list[i].shape)
# save record in txt
mix_name += post_name
crm_name += post_name
# write txt
with open('audio_database/dataset.txt','a') as f:
f.write(mix_name+".npy")
for i in range(len(cRM_list)):
line = " " + crm_name + ("-%05d"%audio_path_list[i][0]) + ".npy"
f.write(line)
f.write("\n")
# save file as npy
np.save(('audio_database/mix/%s.npy'%mix_name), F_mix)
for i in range(len(cRM_list)):
name = crm_name + ("-%05d"%audio_path_list[i][0])
np.save(('audio_database/crm/%s.npy'%name), cRM_list[i])