def split_tr_tt_demand(self):
path_DEMAND = '/media/jeonghwan/HDD2/Dataset/DEMAND/'
save_path = '/media/jeonghwan/HDD2/IS2021/dataset/Simul_DB_ULA4/demand_noise/tr/'
# Channel1을 선택해서 diffuse로 만들었음
noise_list = glob(path_DEMAND + '**/*ch01.wav')
for i, noi in enumerate(noise_list):
aud, fs = librosa.core.load(noi, sr=None, mono=False)
aud = aud[np.newaxis, :]
for j in range(1, 4):
aud_temp, _ = librosa.core.load(noi.replace(
'ch01', 'ch0{}'.format(j + 1)),
sr=None,
mono=False)
aud = np.concatenate((aud, aud_temp[np.newaxis, :]), axis=0)
fn = noi.split('/')[-2] + '_' + noi.split('/')[-1]
len_tr = int(aud.shape[1] * 4 / 5)
noi_tr = aud[:, :len_tr]
noi_tt = aud[:, len_tr:]
audiowrite(save_path + fn.split('/')[-1], noi_tr.T, fs)
audiowrite(
save_path.replace('/tr/', '/tt/') + fn.split('/')[-1],
noi_tt.T, fs)
def mix_rir_and_sound_source_210301(self):
### --------- single source dataset----------- ###
save_path = '/media/jeonghwan/HDD2/IS2021/dataset/SSL/tt/'
rir_path = '/media/jeonghwan/HDD2/IS2021/dataset/simulated_RIR/tr/anechoic/'
spc_path = '/media/jeonghwan/HDD1/Dataset/MS-SNSD/clean_test/'
rir_list = glob(rir_path + '*.npz')
spc_list = glob(spc_path + '*.wav')
# generate random rir index
spc_list.sort()
for i, _spc in enumerate(tqdm(spc_list)):
# read audio file
aud, fs = librosa.core.load(_spc, sr=None, mono=False)
if len(aud.shape) != 1:
aud = aud[:, 0]
#aud.shape[1]
idx_s = np.random.randint(0, len(rir_list))
npz = np.load(rir_list[idx_s], allow_pickle=True)
# convolve
rir = npz['rir']
Y = ss.convolve(rir, aud[:, np.newaxis])
audiowrite(
save_path + rir_list[idx_s].split('/')[-1].split('.')[0] +
'_' + _spc.split('/')[-1], Y, fs)
def convolve_and_save_rir(self, fn):
# read audio file
save_path = '/media/jeonghwan/HDD2/IS2021/dataset/Simul_DB_ULA4/multi_channel_speech/tt/'
rir_path = '/media/jeonghwan/HDD2/IS2021/dataset/Simul_DB_ULA4/simulated_RIR/anechoic/tr/'
mix_all = False
rir_list = glob(rir_path + '*.npz')
aud, fs = librosa.core.load(fn, sr=None, mono=False)
if len(aud.shape) != 1:
aud = aud[:, 0]
if mix_all == True:
for i, _rir in enumerate(rir_list):
npz = np.load(_rir, allow_pickle=True)
rir = npz['rir']
Y = ss.convolve(rir, aud[:, np.newaxis])
audiowrite(
save_path + _rir.split('/')[-1].split('.')[0] + '_' +
fn.split('/')[-1], Y, fs)
else:
idx_s = np.random.randint(0, len(rir_list))
npz = np.load(rir_list[idx_s], allow_pickle=True)
# convolve
rir = npz['rir']
Y = ss.convolve(rir, aud[:, np.newaxis])
audiowrite(
save_path + rir_list[idx_s].split('/')[-1].split('.n')[0] +
'_' + fn.split('/')[-1], Y, fs)
def convolve_and_save_rir_tt(self, fn):
# path set
mode = 'tt'
save_path = os.getcwd() + '/multi_channel_speech/' + mode + '/clean'
Path(save_path).mkdir(parents=True, exist_ok=True)
rir_path = os.getcwd() + '/rir/' + mode
mix_all = False
rir_list = glob(rir_path + '/*/*.npz')
# aud, fs = librosa.core.load(fn, sr=None, mono=False)
aud, fs = audioread(fn)
if len(aud.shape) != 1:
aud = aud[:, 0]
if mix_all == True:
for i, _rir in enumerate(rir_list):
npz = np.load(_rir, allow_pickle=True)
rir = npz['rir']
Y = ss.convolve(rir, aud[:, np.newaxis])
audiowrite(
save_path + '/' + _rir.split('/')[-1].split('.')[0] + '_' +
fn.split('/')[-1], Y, fs)
else:
idx_s = np.random.randint(0, len(rir_list))
npz = np.load(rir_list[idx_s], allow_pickle=True)
# convolve
rir = npz['rir']
Y = ss.convolve(rir, aud[:, np.newaxis])
audiowrite(
save_path + '/' + rir_list[idx_s].split('/')[-2] + '_' +
rir_list[idx_s].split('/')[-1].split('.n')[0] + '_' +
fn.split('/')[-1], Y, fs)
def mix_rir_and_sound_source(self, mode):
"""
convolve speech and speech_rir (random selected)
:param mode: tr/cv/tt
:return: save multi-channel speech
"""
# path set
save_path = os.getcwd() + '/multi_channel_speech/' + mode
rir_path = os.getcwd() + '/rir/' + mode
if mode == 'cv':
rir_path = os.getcwd() + '/rir/tr'
spc_path = '/home/dail/PycharmProjects/DCCRN/data/tr/clean'
# rir list and sound source list
rir_list = glob(rir_path + '/*/*.npz')
spc_list = glob(spc_path + '/*.wav')
# generate random rir index
spc_list.sort()
_use_par = False
if _use_par == True:
if mode == 'tr':
_ = parmap.map(self.convolve_and_save_rir_tr,
spc_list,
pm_pbar=True,
pm_processes=28)
if mode == 'cv':
_ = parmap.map(self.convolve_and_save_rir_cv,
spc_list,
pm_pbar=True,
pm_processes=28)
if mode == 'tt':
_ = parmap.map(self.convolve_and_save_rir_tt,
spc_list,
pm_pbar=True,
pm_processes=28)
else:
for i, _spc in enumerate(tqdm(spc_list)):
# read audio file
# aud, fs = librosa.core.load(_spc, sr=None, mono=False)
aud, fs = audioread(_spc)
if len(aud.shape) != 1:
aud = aud[:, 0]
#aud.shape[1]
idx_s = np.random.randint(0, len(rir_list))
npz = np.load(rir_list[idx_s], allow_pickle=True)
# convolve
rir = npz['rir']
Y = ss.convolve(rir, aud[:, np.newaxis])
audiowrite(
save_path + '/' + rir_list[idx_s].split('/')[-2] + '_' +
rir_list[idx_s].split('/')[-1].split('.n')[0] + '_' +
_spc.split('/')[-1], Y, fs)
def generate_whitenoise(self):
wn = np.random.randn(self.fs * 5,
1) - 0.5 # create random signal [0, 1]
wn = wn / np.max(np.abs(wn)) * 0.5
plt.figure()
plt.plot(wn)
plt.show()
audiowrite('wn.wav', wn, self.fs)
return wn
def mix_spatially_white_noise(self, fn, SNR):
"""
:param fn: audio filename
:param SNR: want to mix
:return:
"""
# add spatially uncorrelated white noise
swn = np.random.randn(Y.shape[0], 4) - 0.5
Y += swn * 0.002
audiowrite('mixed_{}dB.wav'.format(SNR))
return
def convolve_and_save_rir_mp(self, fn):
# read audio file
# save_path = '/media/jeonghwan/HDD2/IS2021/dataset/Simul_DB_ULA4/multi_channel_speech/tr/'
save_path = '/media/jeonghwan/HDD2/IS2021/dataset/Simul_DB_ULA4/multi_channel_speech/tt/'
#save_path = '/media/jeonghwan/HDD2/IS2021/dataset/Simul_DB_ULA4/multi_channel_noisy_direct/tr/'
#save_path = '/media/jeonghwan/HDD2/IS2021/dataset/Simul_DB_ULA4/multi_channel_noisy_direct/tt/'
aud, fs = librosa.core.load(fn[0], sr=None, mono=False)
if len(aud.shape) != 1:
aud = aud[:, 0]
room_num = fn[1].split('/')[-2]
# convolve
npz = np.load(fn[1], allow_pickle=True)
rir = npz['rir']
Y = ss.convolve(rir, aud[:, np.newaxis])
audiowrite(
save_path + fn[1].split('/')[-1].split('.n')[0] + '_' + room_num +
'_' + fn[0].split('/')[-1], Y, fs)
def split_tr_tt_diffuse(self):
path_DEMAND = '/media/jeonghwan/HDD2/Dataset/DEMAND/'
save_path = '/media/jeonghwan/HDD2/IS2021/dataset/Simul_DB_ULA4/diffuse_noise/tr/'
# Channel1을 선택해서 diffuse로 만들었음
noise_list = glob(path_DEMAND + '**/*ch01.wav')
for i, noi in enumerate(noise_list):
aud, fs = librosa.core.load(noi, sr=None, mono=False)
fn = noi.split('/')[-2] + '_' + noi.split('/')[-1]
len_tr = int(aud.shape[0] * 4 / 5)
noi_tr = aud[:len_tr]
noi_tt = aud[len_tr:]
audiowrite(save_path + fn.split('/')[-1], noi_tr, fs)
audiowrite(
save_path.replace('/tr/', '/tt/') + fn.split('/')[-1], noi_tt,
fs)
def generate_single_source(self):
# read audio file
aud, fs = audioread(path='wn.wav', sr=None, mono=False)
# load rir file
npz_path = './'
fn = 'az150_el0_r2.0'
npz_list = glob(npz_path + '{}.npz'.format(fn))
npz = np.load(npz_list[0], allow_pickle=True)
rir = npz['rir']
Y = self.convolve_rir_signal(rir, aud)
# save
audiowrite('{}.wav'.format(fn), Y, 16000)
return Y
def mix_spc_noi_tt(self):
"""
(1) load single channel speech / single channel speech
(2) select speech rir and noise rir
(3) convolve speech(noise) and speech(noise) rir
(4) mix multi-channel speech and multi-channel noise
(5) room(2) * noise(4) * SNR(5)
:return: save noisy(mix), clean(s1), noise(s2) files / save 'output.csv' file
"""
# path set
spc_path = '/home/dail/PycharmProjects/DCCRN/datasets/tr/clean'
noi_path = '/home/dail/PycharmProjects/DCCRN/datasets/tr/noise'
snr_list = [-5, 0, 5, 10, 15]
save_path = os.getcwd() + '/output/tt'
Path(save_path + '/mix').mkdir(parents=True, exist_ok=True)
Path(save_path + '/s1').mkdir(parents=True, exist_ok=True)
Path(save_path + '/s2').mkdir(parents=True, exist_ok=True)
# multi-channel speech list
s_list = glob(spc_path + '/*.wav')
# single-channel noise list
n_list = glob(noi_path + '/*.wav')
# make 'output.csv'
f = open(f'output/tt/output.csv', 'w', newline='')
wr = csv.writer(f)
wr.writerow([
'order', 'speech', 'room', 'speech_rir', 'noise', 'noise_rir',
'snr'
])
cnt = 0
for i, s in enumerate(s_list):
# multi_ch_aud, fs = librosa.core.load(s, sr=None, mono=False)
# multi_ch_aud_na = os.path.splitext(os.path.basename(s))[0]
spc, fs = audioread(s)
spc_na = os.path.splitext(os.path.basename(s))[0]
# select speech/noise rir
# np.random.seed(1)
rand_azi_s = np.random.choice(
np.concatenate((np.arange(31), np.arange(330, 360)), axis=0))
# np.random.seed(1)
rand_azi_n = np.random.choice(np.arange(180, 271))
rand_r = np.round(np.random.choice(np.linspace(1, 2.2, 5)), 1)
spc_rir_na = f'az{rand_azi_s}_el0_r{rand_r}'
noi_rir_na = f'az{rand_azi_n}_el0_r{rand_r}'
room = ['R4', 'R5']
# room
for n in range(2):
spc_rir = os.getcwd() + f'/rir/tt/{room[n]}/{spc_rir_na}.npz'
npz_s = np.load(spc_rir, allow_pickle=True)
rir_s = npz_s['rir']
multi_ch_spc = ss.convolve(rir_s, spc[:, np.newaxis])
multi_ch_spc = multi_ch_spc.transpose()
noi_rir = os.getcwd() + f'/rir/tt/{room[n]}/{noi_rir_na}.npz'
npz_n = np.load(noi_rir, allow_pickle=True)
rir_n = npz_n['rir']
# noise
for idx_n in range(len(n_list)):
noi, fs2 = librosa.core.load(n_list[idx_n], sr=None)
noi_na = os.path.splitext(os.path.basename(
n_list[idx_n]))[0]
assert fs == fs2
rand_start = np.random.randint(
0, noi.shape[0] - multi_ch_spc.shape[1] - 8191)
multi_ch_noi_tmp = ss.convolve(
rir_n, noi[rand_start:rand_start +
multi_ch_spc.shape[1] + 8191, np.newaxis])
multi_ch_noi = multi_ch_noi_tmp[8191:-8191, :].transpose()
# mix speech and noise with SNR
# idx_snr = np.random.randint(0, len(snr_list))
for l in range(len(snr_list)):
cnt = cnt + 1
snr = snr_list[l]
noisy, clean, noise = self.snr_mix(
multi_ch_spc, multi_ch_noi, snr)
audiowrite(
save_path +
f'/mix/noisy_{cnt:#05d}_{noi_na}_{snr}.wav',
noisy.transpose(), fs)
audiowrite(save_path + f'/s1/clean_{cnt:#05d}.wav',
clean.transpose(), fs)
audiowrite(save_path + f'/s2/noise_{cnt:#05d}.wav',
noise.transpose(), fs)
wr.writerow([
cnt, spc_na, room[n], spc_rir_na, noi_na,
noi_rir_na, snr
])
f.close()
def mix_spc_noi_0401(self, mode):
"""
(1) load multi-channel speech
(2) calculate noise rir (=speech rir + 90/180/270 degree)
(3) convolve noise (random selected) and noise rir
(4) mix multi-channel speech and multi-channel noise with SNR (random selected)
:param mode: tr/cv/tt
:return: save noisy(mix), clean(s1), noise(s2) files / save 'output.csv' file
"""
# path set
spc_path = os.getcwd() + '/multi_channel_speech/' + mode
noi_path = os.getcwd() + '/Data/' + mode + '/noise'
snr_list = [-5, 0, 5, 10, 15, 20]
save_path = os.getcwd() + '/output/' + mode
Path(save_path + '/mix').mkdir(parents=True, exist_ok=True)
Path(save_path + '/s1').mkdir(parents=True, exist_ok=True)
Path(save_path + '/s2').mkdir(parents=True, exist_ok=True)
# multi-channel speech list
s_list = glob(spc_path + '/*.wav')
# single-channel noise list
n_list = glob(noi_path + '/*.wav')
# make 'output.csv'
f = open(f'output/{mode}/output.csv', 'w', newline='')
wr = csv.writer(f)
wr.writerow([
'order', 'speech', 'room', 'speech_rir', 'noise', 'noise_rir',
'azimuth', 'snr'
])
for i, s in enumerate(s_list):
multi_ch_aud, fs = librosa.core.load(s, sr=None, mono=False)
multi_ch_aud_na = os.path.splitext(os.path.basename(s))[0]
# select noise azimuth
split = multi_ch_aud_na.split('_')
spc_na = f'{split[-2]}_{split[-1]}'
spc_rir_na = f'{split[1]}_{split[2]}_{split[3]}'
# print(time.time() - start)
# noise rir = speech rir + 90/180/270 degree
az = int(split[1][2:])
room = split[0]
n = np.random.randint(1, 4)
noi_az = (az + 90 * n) % 360 # +90 +180 +270 degree
split = multi_ch_aud_na.split('_')
noi_rir_na = f'az{noi_az}_{split[2]}_{split[3]}'
noi_rir = os.getcwd() + f'/rir/{mode}/{split[0]}/{noi_rir_na}.npz'
if mode == 'cv':
noi_rir = os.getcwd() + f'/rir/tr/{split[0]}/{noi_rir_na}.npz'
# select and load random noise
idx_n = np.random.randint(0, len(n_list))
noi, fs2 = librosa.core.load(n_list[idx_n], sr=None)
noi_na = os.path.splitext(os.path.basename(n_list[idx_n]))[0]
assert fs == fs2
# convolve noise with RIR
npz = np.load(noi_rir, allow_pickle=True)
rir = npz['rir']
rand_start = np.random.randint(
0, noi.shape[0] - multi_ch_aud.shape[1] - 8191)
multi_ch_noi_tmp = ss.convolve(
rir, noi[rand_start:rand_start + multi_ch_aud.shape[1] + 8191,
np.newaxis])
multi_ch_noi = multi_ch_noi_tmp[8191:-8191, :].transpose()
# mix speech and noise with SNR
idx_snr = np.random.randint(0, len(snr_list))
snr = snr_list[idx_snr]
noisy, clean, noise = self.snr_mix(multi_ch_aud, multi_ch_noi, snr)
audiowrite(
save_path + f'/mix/noisy_{i + 1:#05d}_{noi_na}_{snr}.wav',
noisy.transpose(), fs)
audiowrite(save_path + f'/s1/clean_{i + 1:#05d}.wav',
clean.transpose(), fs)
audiowrite(save_path + f'/s2/noise_{i + 1:#05d}.wav',
noise.transpose(), fs)
wr.writerow(
[i, spc_na, room, spc_rir_na, noi_na, noi_rir_na, 90 * n, snr])
f.close()
