for i in range(0, n_frames_mix, mvdr_step):
for j in range(min(mvdr_step, n_frames_mix - i - 1)):
psd_curr = np.zeros((257, 66, 66), dtype=np.complex)
for k in range(n_bins):
psd_curr[k] = np.outer(stft_mix_noise[k, :, i + j],
stft_mix_noise[k, :, i + j].conj())
psd_curr *= taper
psd = EXP_AV_COEF * psd + (1 - EXP_AV_COEF) * psd_curr
psd_reg = psd + REG_COEF * np.identity(psd.shape[-1])
w = get_mvdr_vector(d_arr.T, psd_reg)
if i == 0:
taper, _ = cov_matrix_tapper_bandwidth(stft_mix[:, :, 0:mvdr_step],
hor_mic_count,
vert_mic_count,
dHor=dHor,
dVert=dVert,
angle_v=angle_v,
angle_h=angle_h,
sr=sr)
w = np.zeros((n_bins, n_sensors))
pp = apply_beamforming_vector(w, stft_mix[:, :, i:i + mvdr_step])
res_spec = pp
w = np.zeros((n_bins, n_sensors))
#################################################################
# 4 - Calc psd matrix
psd_noise_matrix = get_power_spectral_density_matrix(stft_noise_arr)
print ('Calc psd matrix done!')
print (' psd_noise_matrix.shape = ', psd_noise_matrix.shape)
np.save('psd_noise_mat', psd_noise_matrix)
#################################################################
# 5 - Regularisation psd matrix
psd_noise_matrix = psd_noise_matrix + 0.001*np.identity(psd_noise_matrix.shape[-1])
#################################################################
# 6 - Apply MVDR
w = get_mvdr_vector(d_arr.T, psd_noise_matrix)
result_mvdr_spec = apply_beamforming_vector(w, stft_mix_arr)
print ('Apply MVDR done!')
#################################################################
# 7 - Do align
align_stft_arr = ds_align(stft_mix_arr, d_arr.T)
print ('Align mix_arr done!')
#################################################################
# 8 - Calc mccowan filter output
# 8.1 - Get coherence shape - (sensors_count, sensors_count, bins)
G = diffuse_noise_coherence(bins = n_bins, freq = sr, D_IJ = pair_distance)
# 8.2 - Calc mccowan filter output
