def single_normal():
# audio_data = get_audio_nochime('data/new_dataset/216m/2m_pub_new', ch_range=range(1, 9), fs=16000)
# noise_data = get_audio_nochime('data/new_dataset/blstm_noise/noise_124', ch_range=range(1, 9), fs=16000)
# audio_data = get_audio_nochime(args.data_directory, ch_range=range(1, 3), fs=16000)
t_io = 0
t_net = 0
t_beamform = 0
# check execution time
with Timer() as t:
audio_data = get_audio_nochime(args.data_directory,
ch_range=range(1, 3),
fs=16000)
context_samples = 0
print("audio_data: ", audio_data.shape, end="\n")
# for i in range (0, 8):
# print(audio_data[i][1])
t_io += t.msecs
Y = stft(audio_data, time_dim=1).transpose((1, 0, 2))
# N = stft(noise_data, time_dim=1).transpose((1, 0, 2))
Y_phase = np.divide(Y, abs(Y))
print("Y: ", Y.shape, "Y_phase: ", Y_phase.shape, end="\n")
# Y_var with or without chainer Variable class doesn't give any different
Y_var = Variable(np.abs(Y).astype(np.float32))
# N_var = Variable(np.abs(N).astype(np.float32), True)
# blstm_noise = Variable(np.abs(blstm_noise).astype(np.float32), True)
with Timer() as t:
# mask estimation
N_masks, X_masks = model.calc_masks(Y_var)
# Noise_masks = model.calc_mask_noise(N_var)
print("N_masks: ", N_masks.shape, end="\n")
N_masks.to_cpu()
X_masks.to_cpu()
t_net += t.msecs
# Noise_masks.to_cpu()
with Timer() as t:
N_mask = np.median(N_masks.data, axis=1)
X_mask = np.median(X_masks.data, axis=1)
# Noise_mask = np.median(Noise_masks.data, axis=1)
# signal = audioread('data/new_dataset/216m/2m_pub_new' + '.CH{}.wav'.format(ch), sample_rate=16000)
# noise = audioread('data/new_dataset/gevnoise/gevnoise' + '.CH{}.wav'.format(ch), sample_rate=16000)
# signal_ = stft(signal)
# noise_ = stft(noise)
#
# signal_phase = np.divide(signal, abs(signal_))
# noise_masks = model.calc_mask_noise(noise_)
# noise_to = np.multiply(noise_masks.data, signal_)
# noise_to = np.multiply(noise_to, signal_phase)
# audiowrite(istft(noise_to)[context_samples:],
# "/home/hipo/workspace/BeamSaber/result/noise/noise_to_.CH{}.wav".format(ch), 16000, True, True)
Noise = np.multiply(N_masks.data, Y)
Noise = np.multiply(Noise, Y_phase)
# Y_phase_med = np.median(Y_phase, axis=1)
# print(Noise.shape)
# for ch in range(0, 8):
# audiowrite(istft(Noise[:,ch,:])[context_samples:],
# "/home/hipo/workspace/BeamSaber/result/noise/2mnoise_.CH{}.wav".format(ch), 16000, True, True)
Noise = np.median(Noise, axis=1)
# print("N_mask: ", N_mask.shape, "X_mask: ", X_mask.shape, "Y_phase: ", Y_phase.shape, end="\n")
Y_hat = gev_wrapper_on_masks(Y, N_mask, X_mask)
# print(Y_hat.shape)
# print("Noise: ", Noise.shape)
t_beamform += t.msecs
with Timer() as t:
audiowrite(
istft(Noise)[context_samples:],
"/media/hipo/lento/workspace/BeamSaber/tools/enhancement/gev/PublicFOMLSA/sample/{}_noise.wav"
.format(args.exNum), 16000, True, True)
audiowrite(
istft(Y_hat)[context_samples:],
"/media/hipo/lento/workspace/BeamSaber/tools/enhancement/gev/PublicFOMLSA/sample/{}_gev.wav"
.format(args.exNum), 16000, True, True)
t_io += t.msecs
print(
'Timings: I/O: {:.2f}s | Net: {:.2f}s | Beamformer: {:.2f}s | Total: {:.2f}s'
.format(t_io / 1000, t_net / 1000, t_beamform / 1000,
((t_io + t_net + t_beamform) / 1000)))
cur_line[0], cur_line[1], cur_line[2])
t_io += t.msecs
Y = stft(audio_data, time_dim=1).transpose((1, 0, 2))
Y_var = Variable(np.abs(Y).astype(np.float32))
if args.gpu >= 0:
Y_var.to_gpu(args.gpu)
with Timer() as t:
N_masks, X_masks = model.calc_masks(Y_var)
N_masks.to_cpu()
X_masks.to_cpu()
t_net += t.msecs
with Timer() as t:
N_mask = np.median(N_masks.data, axis=1)
X_mask = np.median(X_masks.data, axis=1)
Y_hat = gev_wrapper_on_masks(Y, N_mask, X_mask)
t_beamform += t.msecs
if scenario == 'simu':
wsj_name = cur_line.split('/')[-1].split('_')[1]
spk = cur_line.split('/')[-1].split('_')[0]
env = cur_line.split('/')[-1].split('_')[-1]
elif scenario == 'real':
wsj_name = cur_line[3]
spk = cur_line[0].split('/')[-1].split('_')[0]
env = cur_line[0].split('/')[-1].split('_')[-1]
filename = os.path.join(args.output_dir,
'{}05_{}_{}'.format(stage, env.lower(), scenario),
'{}_{}_{}.wav'.format(spk, wsj_name, env.upper()))
cur_line[0], cur_line[1], cur_line[2])
t_io += t.msecs
Y = stft(audio_data, time_dim=1).transpose((1, 0, 2))
Y_var = Variable(np.abs(Y).astype(np.float32), True)
if args.gpu >= 0:
Y_var.to_gpu(args.gpu)
with Timer() as t:
N_masks, X_masks = model.calc_masks(Y_var)
N_masks.to_cpu()
X_masks.to_cpu()
t_net += t.msecs
with Timer() as t:
N_mask = np.median(N_masks.data, axis=1)
X_mask = np.median(X_masks.data, axis=1)
Y_hat = gev_wrapper_on_masks(Y, N_mask, X_mask)
t_beamform += t.msecs
if scenario == 'simu':
wsj_name = cur_line.split('/')[-1].split('_')[1]
spk = cur_line.split('/')[-1].split('_')[0]
env = cur_line.split('/')[-1].split('_')[-1]
elif scenario == 'real':
wsj_name = cur_line[3]
spk = cur_line[0].split('/')[-1].split('_')[0]
env = cur_line[0].split('/')[-1].split('_')[-1]
filename = os.path.join(
args.output_dir,
'{}05_{}_{}'.format(stage, env.lower(), scenario),
'{}_{}_{}.wav'.format(spk, wsj_name, env.upper())
N_masks, X_masks = model.calc_masks(Y_var)
N_masks.to_cpu()
X_masks.to_cpu()
t_net += t.msecs
with Timer() as t:
N_mask = np.median(N_masks.data, axis=1)
X_mask = np.median(X_masks.data, axis=1)
print("Y: ",
Y.shape,
"N_mask: ",
N_mask.shape,
"X_mask: ",
X_mask.shape,
end="\n")
Y_hat = gev_wrapper_on_masks(Y, N_mask, X_mask)
# audiowrite(istft(Y_hat), "new_dataset_result/2m_feedback_.wav", 48000, True, True)
t_beamform += t.msecs
# second pass beamforming
# second_channel = audioread('AUDIO_RECORDING.CH2.wav', sample_rate=48000)
second_channel = audioread('new_dataset/2m/2m_pub_new.CH5.wav',
sample_rate=48000)
second_channel = np.expand_dims(second_channel, axis=0)
print("second_size", second_channel.shape, end="\n")
second_channel = stft(second_channel, time_dim=1).transpose((1, 0, 2))
print("Y_hat: ", Y_hat.shape, "second_size", second_channel.shape, end="\n")
Y_hat = np.expand_dims(Y_hat, axis=1)
Y_var_second = Variable(np.abs(Y_hat).astype(np.float32), True)
t_io += t.msecs
Y = stft(audio_data, time_dim=1).transpose((1, 0, 2))
Y_var = Variable(np.abs(Y).astype(np.float32), True)
if args.gpu >= 0:
Y_var.to_gpu(args.gpu)
with Timer() as t:
N_masks, X_masks = model.calc_masks(Y_var)
N_masks.to_cpu()
X_masks.to_cpu()
t_net += t.msecs
with Timer() as t:
data_tmp = X_masks.data
N_mask = np.median(N_masks.data, axis=1)
X_mask = np.median(X_masks.data, axis=1)
Y_hat = gev_wrapper_on_masks(Y, N_mask, X_mask, output_setup,
corr_info)
#N_mask = N_masks.data
#X_mask = X_masks.data
#Y_hat = mcmf_wrapper_on_masks(Y, N_mask, X_mask, output_setup, corr_info)
t_beamform += t.msecs
# the spliter in Win '\' and Linux '/'
if scenario == 'simu':
wsj_name = cur_line.split('\\')[-1].split('_')[1]
spk = cur_line.split('\\')[-1].split('_')[0]
env = cur_line.split('\\')[-1].split('_')[-1]
elif scenario == 'real':
wsj_name = cur_line[3]
spk = cur_line[0].split('\\')[-1].split('_')[0]
env = cur_line[0].split('\\')[-1].split('_')[-1]
