# Geometry of the room and location of sources and microphones
room_dim = np.array([10, 7.5, 3])
mic_locs = np.vstack((
pra.circular_2D_array([4.1, 3.76], n_mics, np.pi / 2, 0.02),
1.2 * np.ones((1, n_mics)),
))
target_locs = semi_circle_layout([4.1, 3.755, 1.1],
np.pi / 2,
2.0,
n_sources_target,
rot=0.743 * np.pi)
# interferer_locs = grid_layout([3., 5.5], n_sources - n_sources_target, offset=[6.5, 1., 1.7])
interferer_locs = random_layout([3.0, 5.5, 1.5],
n_sources - n_sources_target,
offset=[6.5, 1.0, 0.5],
seed=1)
source_locs = np.concatenate((target_locs, interferer_locs), axis=1)
# Prepare the signals
wav_files = sampling(1,
n_sources,
f"{samples_dir}/metadata.json",
gender_balanced=True,
seed=8)[0]
signals = wav_read_center(wav_files, seed=123)
# Place the blinkies regularly in the room (2D plane)
blinky_locs = gm_layout(n_blinkies,
target_locs - np.c_[[0.0, 0.0, 0.4]],
std=[0.4, 0.4, 0.05],
def createroom(mic_p, mic_d, sour_p, sour_d, callback_mix, roomdim,
absorption, max_order, n_mics, angle):
np.random.seed(10)
# STFT parameters
framesize = 4096
win_a = pra.hann(framesize)
win_s = pra.transform.compute_synthesis_window(win_a, framesize // 2)
# algorithm parameters
# param ogive
ogive_mu = 0.1
ogive_update = "switching"
ogive_iter = 2000
SIR = 10 # dB
SNR = (
60
) # dB, this is the SNR with respect to a single target source and microphone self-noise
########separation params#############
algo = algo_choices[0]
no_cb = True
save = True
n_iter = 60
dist = "gauss" #guass or laplace
########paramas set##################
fs = 44100
n_sources = 2
n_mics = n_mics
n_sources_target = 2
assert n_sources_target <= n_mics, "More sources than microphones is not supported"
# set the source powers, the first one is half
source_std = np.ones(n_sources_target)
# room size
room_dim = roomdim
# micro position
rot = angle
offset = np.pi - rot / 2
mic_locs = semi_circle_layout(mic_p, rot, mic_d, n_mics,
rot=offset) ###micro2
# target position
target_locs = np.transpose([[7, 10, 6], [9, 16, 6]])
#interference position
interferer_locs = random_layout([14, 0, 6],
n_sources - n_sources_target,
offset=[5, 20, 3],
seed=1)
source_locs = target_locs
# audio loaded
wav_files = [amBird, saBird]
signals = wav_read_center(wav_files, seed=123)
#create room
room = pra.ShoeBox(room_dim,
fs=44100,
absorption=absorption,
max_order=max_order,
air_absorption=True,
humidity=50)
# add source
for sig, loc in zip(signals, source_locs.T):
room.add_source(loc, signal=sig)
# add micro
room.add_microphone_array(pra.MicrophoneArray(mic_locs, fs=room.fs))
callback_mix_kwargs = {
"snr": SNR,
"sir": SIR,
"n_src": n_sources,
"n_tgt": n_sources_target,
"src_std": source_std,
"ref_mic": 0,
}
# Run the simulation
separate_recordings = room.simulate(
callback_mix=callback_mix,
callback_mix_kwargs=callback_mix_kwargs,
return_premix=True,
)
mics_signals = room.mic_array.signals
print("Simulation done.")
# rt60 = room.measure_rt60()
# print(rt60)
# Monitor Convergence
ref = np.moveaxis(separate_recordings, 1, 2)
if ref.shape[0] < n_mics:
ref = np.concatenate(
(ref,
np.random.randn(n_mics - ref.shape[0], ref.shape[1],
ref.shape[2])),
axis=0,
)
SDR, SIR, cost_func = [], [], []
convergence_callback = None
# START BSS
# shape: (n_frames, n_freq, n_mics)
X_all = pra.transform.analysis(mics_signals.T,
framesize,
framesize // 2,
win=win_a).astype(np.complex128)
X_mics = X_all[:, :, :n_mics]
tic = time.perf_counter()
# Run BSS
if algo == "auxiva":
# Run AuxIVA
Y = overiva(
X_mics,
n_iter=n_iter,
proj_back=True,
model=dist,
callback=convergence_callback,
)
elif algo == "auxiva_pca":
# Run AuxIVA
Y = auxiva_pca(
X_mics,
n_src=n_sources_target,
n_iter=n_iter,
proj_back=True,
model=dist,
callback=convergence_callback,
)
elif algo == "overiva":
# Run AuxIVA
Y = overiva(
X_mics,
n_src=n_sources_target,
n_iter=n_iter,
proj_back=True,
model=dist,
init_eig=(init == init_choices[1]),
callback=convergence_callback,
)
elif algo == "ilrma":
# Run AuxIVA
Y = pra.bss.ilrma(
X_mics,
n_iter=n_iter,
n_components=2,
proj_back=True,
callback=convergence_callback,
)
elif algo == "ogive":
# Run OGIVE
Y = ogive(
X_mics,
n_iter=ogive_iter,
step_size=ogive_mu,
update=ogive_update,
proj_back=True,
model=dist,
init_eig=(init == init_choices[1]),
callback=convergence_callback,
)
elif algo == "ogive_matlab":
# Run OGIVE
Y = ogive_matlab_wrapper(
X_mics,
n_iter=ogive_iter,
step_size=ogive_mu,
update=ogive_update,
proj_back=True,
init_eig=(init == init_choices[1]),
callback=convergence_callback,
)
else:
raise ValueError("No such algorithm {}".format(algo))
toc = time.perf_counter()
# Run iSTFT
if Y.shape[2] == 1:
y = pra.transform.synthesis(Y[:, :, 0],
framesize,
framesize // 2,
win=win_s)[:, None]
y = y.astype(np.float64)
else:
y = pra.transform.synthesis(Y,
framesize,
framesize // 2,
win=win_s).astype(np.float64)
# If some of the output are uniformly zero, just add a bit of noise to compare
for k in range(y.shape[1]):
if np.sum(np.abs(y[:, k])) < 1e-10:
y[:, k] = np.random.randn(y.shape[0]) * 1e-10
# For conventional methods of BSS, reorder the signals by decreasing power
if algo != "blinkiva":
new_ord = np.argsort(np.std(y, axis=0))[::-1]
y = y[:, new_ord]
# Compare SIR
m = np.minimum(y.shape[0] - framesize // 2, ref.shape[1])
sdr, sir, sar, perm = bss_eval_sources(
ref[:n_sources_target, :m, 0],
y[framesize // 2:m + framesize // 2, :n_sources_target].T,
)
# reorder the vector of reconstructed signals
y_hat = y[:, perm]
print("SDR:", sdr)
print("SIR:", sir)
####save mix and separation #######
if save:
from scipy.io import wavfile
wavfile.write(
"birdmix.wav",
room.fs,
(pra.normalize(mics_signals, bits=16).astype(np.int16).T)[:,
0],
)
for i, sig in enumerate(y_hat.T):
wavfile.write(
"birdsep{}.wav".format(i + 1),
room.fs,
pra.normalize(sig, bits=16).astype(np.int16).T,
)
def createroom(amBird, saBird, noises, mic_p, mic_d, sour_p, sour_d, callback_mix, roomdim, absorption, max_order, n_mics, angle):
###########IVA params set################
np.random.seed(10)
# STFT parameters
framesize = 4096
win_a = pra.hann(framesize)
win_s = pra.transform.compute_synthesis_window(
win_a, framesize // 2)
# algorithm parameters
# param ogive
ogive_mu = 0.1
ogive_update = "switching"
ogive_iter = 2000
# dB, this is the SNR with respect to a single target source and microphone self-noise
########separation params##################
algo = algo_choices[0]
no_cb = True
save = True
n_iter = 60
dist = "gauss" # guass or laplace
########paramas set##################
fs = 44100
snr=60
sinr=10
n_sources = 2+3
n_mics = n_mics
n_sources_target = 2
assert n_sources_target <= n_mics, "More sources than microphones is not supported"
# set the source powers, the first one is half
source_std = np.ones(n_sources_target)
# position
# room size
room_dim = roomdim
# micro position
rot = angle
offset = np.pi-rot/2
mic_locs = semi_circle_layout(
mic_p, rot, mic_d, n_mics, rot=offset) # micro2
# target position
target_locs = np.transpose([[7, 10, 6], [9, 16, 6]])
#interfere position
interferer_locs = random_layout([16, 2, 6], 3, offset=[5, 18, 3], seed=1)
source_locs = np.concatenate((target_locs, interferer_locs), axis=1)
# source_locs = target_locs
# audio input
wav_files = [amBird, saBird, noises[0],noises[1],noises[2]]
signals = wav_read_center(wav_files, seed=123)
# create room
room = pra.ShoeBox(room_dim, fs=44100, absorption=absorption,
max_order=max_order, air_absorption=True, humidity=50)
# add source
for sig, loc in zip(signals, source_locs.T):
room.add_source(loc, signal=sig)
# add micro
room.add_microphone_array(
pra.MicrophoneArray(mic_locs, fs=room.fs))
# # draw
# x = mic_locs[:2][0]
# y = mic_locs[:2][1]
# import matplotlib.pyplot as plt
# plt.scatter(x,y)
# plt.axis('equal')
# plt.xlim([0,20])
# plt.ylim([0,20])
# x1 = source_locs[:2][0]
# y1 = source_locs[:2][1]
# plt.scatter(x1,y1)
# plt.xlim([0,20])
# plt.ylim([0,20])
# plt.axis('equal')
# x1 = interferer_locs[:2][0]
# y1 = interferer_locs[:2][1]
# plt.scatter(x1,y1)
# plt.xlim([0,20])
# plt.ylim([0,20])
# plt.axis('equal')
# plt.show()
# callback_mix_kwargs = {
# "snr": SNR,
# "sir": SIR,
# "n_src": n_sources,
# "n_tgt": n_sources_target,
# "src_std": source_std,
# "ref_mic": 0,
# }
# power set
premix = room.simulate(return_premix=True)
n_samples = premix.shape[2]
# Normalize the signals so that they all have unit variance at the reference microphone
ref_mic=0
p_mic_ref = np.std(premix[:, ref_mic, :], axis=1)
premix /= p_mic_ref[:, None, None]
sources_var = np.ones(n_sources_target) #
# scale to pre-defined variance
premix[:n_sources_target, :, :] *= np.sqrt(sources_var[:, None, None])
# compute noise variance
sigma_n = np.sqrt(10 ** (-snr / 10) * np.sum(sources_var))
# now compute the power of interference signal needed to achieve desired SINR
sigma_i = np.sqrt(
np.maximum(0, 10 ** (-sinr / 10) * np.sum(sources_var) - sigma_n ** 2)
/ (n_sources-n_sources_target)
)
premix[n_sources_target:, :, :] *= sigma_i
background = (
np.sum(premix[n_sources_target:, :, :], axis=0)
)
# Mix down the recorded signals
mix = np.sum(premix, axis=0)
mics_signals = room.mic_array.signals
print("Simulation done.")
# Monitor Convergence
ref = np.zeros(
(n_sources_target+1, premix.shape[2], premix.shape[1]), dtype=premix.dtype)
ref[:n_sources_target, :, :] = premix[:n_sources_target, :, :].swapaxes(1, 2)
ref[n_sources_target, :, :] = background.T
convergence_callback = None
# START BSS
X_all = pra.transform.analysis(
mix.T, framesize, framesize // 2, win=win_a
).astype(np.complex128)
X_mics = X_all[:, :, :n_mics]
# Run BSS
if algo == "auxiva":
# Run AuxIVA
Y = overiva(
X_mics,
n_iter=n_iter,
proj_back=True,
model=dist,
callback=convergence_callback,
)
elif algo == "auxiva_pca":
# Run AuxIVA
Y = auxiva_pca(
X_mics,
n_src=n_sources_target,
n_iter=n_iter,
proj_back=True,
model=dist,
callback=convergence_callback,
)
elif algo == "overiva":
# Run AuxIVA
Y = overiva(
X_mics,
n_src=n_sources_target,
n_iter=n_iter,
proj_back=True,
model=dist,
init_eig=(init == init_choices[1]),
callback=convergence_callback,
)
elif algo == "ilrma":
# Run AuxIVA
Y = pra.bss.ilrma(
X_mics,
n_iter=n_iter,
n_components=2,
proj_back=True,
callback=convergence_callback,
)
elif algo == "ogive":
# Run OGIVE
Y = ogive(
X_mics,
n_iter=ogive_iter,
step_size=ogive_mu,
update=ogive_update,
proj_back=True,
model=dist,
init_eig=(init == init_choices[1]),
callback=convergence_callback,
)
elif algo == "ogive_matlab":
# Run OGIVE
Y = ogive_matlab_wrapper(
X_mics,
n_iter=ogive_iter,
step_size=ogive_mu,
update=ogive_update,
proj_back=True,
init_eig=(init == init_choices[1]),
callback=convergence_callback,
)
else:
raise ValueError("No such algorithm {}".format(algo))
# Run iSTFT
if Y.shape[2] == 1:
y = pra.transform.synthesis(Y[:, :, 0], framesize, framesize // 2, win=win_s)[
:, None
]
y = y.astype(np.float64)
else:
y = pra.transform.synthesis(Y, framesize, framesize // 2, win=win_s).astype(
np.float64
)
# If some of the output are uniformly zero, just add a bit of noise to compare
for k in range(y.shape[1]):
if np.sum(np.abs(y[:, k])) < 1e-10:
y[:, k] = np.random.randn(y.shape[0]) * 1e-10
# For conventional methods of BSS, reorder the signals by decreasing power
if algo != "blinkiva":
new_ord = np.argsort(np.std(y, axis=0))[::-1]
y = y[:, new_ord]
# Compare SIR
m = np.minimum(y.shape[0] - framesize // 2, ref.shape[1])
sdr, sir, sar, perm = bss_eval_sources(
ref[:n_sources_target, :m, 0],
y[framesize // 2: m + framesize // 2, :n_sources_target].T,
)
# reorder the vector of reconstructed signals
y_hat = y[:, perm]
SDRList.append(sdr)
SIRList.append(sir)
print("SDR:", sdr)
print("SIR:", sir)
####save mix and separation #######
if save:
from scipy.io import wavfile
wavfile.write(
"birdmix.wav",
room.fs,
(pra.normalize(mics_signals, bits=16).astype(
np.int16).T)[:, 0],
)
for i, sig in enumerate(y_hat.T):
wavfile.write(
"birdsep{}.wav".format(i + 1),
room.fs,
pra.normalize(sig, bits=16).astype(np.int16).T,
)
def plot_room_setup(filename, n_mics, n_targets, parameters):
'''
Plot the room scenario in 2D
'''
n_interferers = parameters['n_interferers']
n_blinkies = parameters['n_blinkies']
ref_mic = parameters['ref_mic']
room_dim = np.array(parameters['room_dim'])
# total number of sources
n_sources = n_interferers + n_targets
# Geometry of the room and location of sources and microphones
interferer_locs = random_layout([3., 5.5, 1.5],
n_interferers,
offset=[6.5, 1., 0.5],
seed=1)
target_locs = semi_circle_layout(
[4.1, 3.755, 1.2],
np.pi / 1.5,
2., # 120 degrees arc, 2 meters away
n_targets,
rot=0.743 * np.pi,
)
source_locs = np.concatenate((target_locs, interferer_locs), axis=1)
if parameters['blinky_geometry'] == 'gm':
''' Normally distributed in the vicinity of each source '''
blinky_locs = gm_layout(
n_blinkies,
target_locs - np.c_[[0., 0., 0.5]],
std=[0.4, 0.4, 0.05],
seed=987,
)
elif parameters['blinky_geometry'] == 'grid':
''' Placed on a regular grid, with a little bit of noise added '''
blinky_locs = grid_layout([3., 5.5],
n_blinkies,
offset=[1., 1., 0.7],
seed=987)
else:
''' default is semi-circular '''
blinky_locs = semi_circle_layout(
[4.1, 3.755, 1.1],
np.pi,
3.5,
n_blinkies,
rot=0.743 * np.pi - np.pi / 4,
seed=987,
)
mic_locs = np.vstack((
pra.circular_2D_array([4.1, 3.76], n_mics, np.pi / 2, 0.02),
1.2 * np.ones((1, n_mics)),
))
all_locs = np.concatenate((mic_locs, blinky_locs), axis=1)
# Create the room itself
room = pra.ShoeBox(room_dim[:2])
for loc in source_locs.T:
room.add_source(loc[:2])
# Place the microphone array
room.add_microphone_array(pra.MicrophoneArray(all_locs[:2, :], fs=room.fs))
room.plot(img_order=0)
plt.xlim([-0.1, room_dim[0] + 0.1])
plt.ylim([-0.1, room_dim[1] + 0.1])
plt.savefig(filename)
def one_loop(args):
import numpy
np = numpy
import pyroomacoustics
pra = pyroomacoustics
import sys
sys.path.append(parameters['base_dir'])
from routines import semi_circle_layout, random_layout, gm_layout, grid_layout
from blinkiva import blinkiva
from blinkiva_gauss import blinkiva_gauss
from generate_samples import wav_read_center
n_targets, n_mics, rt60, sinr, wav_files, seed = args
# this is the underdetermined case. We don't do that.
if n_mics < n_targets:
return []
# set MKL to only use one thread if present
try:
import mkl
mkl.set_num_threads(1)
except ImportError:
pass
# set the RNG seed
rng_state = np.random.get_state()
np.random.seed(seed)
# STFT parameters
framesize = parameters['stft_params']['framesize']
win_a = pra.hann(framesize)
win_s = pra.transform.compute_synthesis_window(win_a, framesize // 2)
# Generate the audio signals
# get the simulation parameters from the json file
# Simulation parameters
n_repeat = parameters['n_repeat']
fs = parameters['fs']
snr = parameters['snr']
n_interferers = parameters['n_interferers']
n_blinkies = parameters['n_blinkies']
ref_mic = parameters['ref_mic']
room_dim = np.array(parameters['room_dim'])
sources_var = np.ones(n_targets)
sources_var[0] = parameters['weak_source_var']
# total number of sources
n_sources = n_interferers + n_targets
# Geometry of the room and location of sources and microphones
interferer_locs = random_layout([3., 5.5, 1.5], n_interferers, offset=[6.5, 1., 0.5], seed=1)
target_locs = semi_circle_layout(
[4.1, 3.755, 1.2],
np.pi / 1.5, 2., # 120 degrees arc, 2 meters away
n_targets,
rot=0.743 * np.pi,
)
source_locs = np.concatenate((target_locs, interferer_locs), axis=1)
if parameters['blinky_geometry'] == 'gm':
''' Normally distributed in the vicinity of each source '''
blinky_locs = gm_layout(
n_blinkies, target_locs - np.c_[[0., 0., 0.5]],
std=[0.4, 0.4, 0.05], seed=987,
)
elif parameters['blinky_geometry'] == 'grid':
''' Placed on a regular grid, with a little bit of noise added '''
blinky_locs = grid_layout([3.,5.5], n_blinkies, offset=[1., 1., 0.7], seed=987,)
else:
''' default is semi-circular '''
blinky_locs = semi_circle_layout(
[4.1, 3.755, 1.1],
np.pi, 3.5,
n_blinkies,
rot=0.743 * np.pi - np.pi / 4,
seed=987,
)
mic_locs = np.vstack((
pra.circular_2D_array([4.1, 3.76], n_mics, np.pi / 2, 0.02),
1.2 * np.ones((1, n_mics)),
))
all_locs = np.concatenate((mic_locs, blinky_locs), axis=1)
signals = wav_read_center(wav_files, seed=123)
# Create the room itself
room = pra.ShoeBox(room_dim,
fs=fs,
absorption=parameters['rt60_list'][rt60]['absorption'],
max_order=parameters['rt60_list'][rt60]['max_order'],
)
# Place all the sound sources
for sig, loc in zip(signals[-n_sources:,:], source_locs.T,):
room.add_source(loc, signal=sig)
assert len(room.sources) == n_sources, 'Number of signals ({}) doesn''t match number of sources ({})'.format(signals.shape[0], n_sources)
# Place the microphone array
room.add_microphone_array(
pra.MicrophoneArray(all_locs, fs=room.fs)
)
# compute RIRs
room.compute_rir()
# Run the simulation
premix = room.simulate(return_premix=True)
# Normalize the signals so that they all have unit
# variance at the reference microphone
p_mic_ref = np.std(premix[:,ref_mic,:], axis=1)
premix /= p_mic_ref[:,None,None]
# scale to pre-defined variance
premix[:n_targets,:,:] *= np.sqrt(sources_var[:,None,None])
# compute noise variance
sigma_n = np.sqrt(10 ** (- snr / 10) * np.sum(sources_var))
# now compute the power of interference signal needed to achieve desired SINR
sigma_i = np.sqrt(
np.maximum(0, 10 ** (- sinr / 10) * np.sum(sources_var) - sigma_n ** 2) / n_interferers
)
premix[n_targets:,:,:] *= sigma_i
# Mix down the recorded signals
mix = np.sum(premix, axis=0) + sigma_n * np.random.randn(*premix.shape[1:])
ref = np.moveaxis(premix, 1, 2)
# START BSS
###########
# pre-emphasis on blinky signals
if parameters['use_pre_emphasis']:
mix[n_mics:,:-1] = np.diff(mix[n_mics:,:], axis=1)
mix[n_mics:,-1] = 0.
# shape: (n_frames, n_freq, n_mics)
X_all = pra.transform.analysis(
mix.T,
framesize, framesize // 2,
win=win_a,
)
X_mics = X_all[:,:,:n_mics]
U_blinky = np.sum(np.abs(X_all[:,:,n_mics:]) ** 2, axis=1) # shape: (n_frames, n_blinkies)
# convergence monitoring callback
def convergence_callback(Y, n_targets, SDR, SIR, ref, framesize, win_s, algo_name):
from mir_eval.separation import bss_eval_sources
y = pra.transform.synthesis(Y, framesize, framesize // 2, win=win_s,)
if not algo_name.startswith('blinkiva'):
new_ord = np.argsort(np.std(y, axis=0))[::-1]
y = y[:,new_ord]
m = np.minimum(y.shape[0]-framesize//2, ref.shape[1])
sdr, sir, sar, perm = bss_eval_sources(
ref[:n_targets,:m,0],
y[framesize//2:m+framesize//2,:n_targets].T,
)
SDR.append(sdr.tolist())
SIR.append(sir.tolist())
# store results in a list, one entry per algorithm
results = []
for name, kwargs in parameters['algorithm_kwargs'].items():
results.append({
'algorithm' : name,
'n_targets' : n_targets,
'n_mics' : n_mics,
'rt60' : rt60,
'sinr' : sinr,
'seed' : seed,
'sdr' : [], 'sir' : [], # to store the result
})
if parameters['monitor_convergence']:
cb = lambda Y : convergence_callback(
Y, n_targets,
results[-1]['sdr'], results[-1]['sir'],
ref, framesize, win_s, name,
)
else:
cb = None
# In that case, we still want to capture the initial values of SDR/SIR
convergence_callback(
X_mics, n_targets,
results[-1]['sdr'], results[-1]['sir'],
ref, framesize, win_s, name,
)
if name == 'auxiva':
# Run AuxIVA
Y = pra.bss.auxiva(
X_mics,
callback=cb,
**kwargs,
)
elif name == 'blinkiva':
# Run BlinkIVA
Y = blinkiva(
X_mics, U_blinky, n_src=n_targets,
callback=cb,
**kwargs,
)
elif name == 'blinkiva-gauss':
# Run BlinkIVA
Y = blinkiva_gauss(
X_mics, U_blinky, n_src=n_targets,
callback=cb,
**kwargs,
)
else:
continue
# The last evaluation
convergence_callback(
Y, n_targets,
results[-1]['sdr'], results[-1]['sir'],
ref, framesize, win_s, name,
)
# restore RNG former state
np.random.set_state(rng_state)
return results
