def no_overlap(D):
if D == 1:
x_local = x[:, 0]
else:
x_local = x[:, :D]
hop = block_size
# analysis
X = stft.analysis(x_local, L=block_size, hop=hop)
# synthesis
x_r = stft.synthesis(X, L=block_size, hop=hop)
return pra.dB(np.max(np.abs(x_local - x_r)))
def half_overlap(D):
if D == 1:
x_local = x[:, 0]
else:
x_local = x[:, :D]
hop = block_size // 2
# analysis
analysis_win = pra.hann(block_size)
X = stft.analysis(x_local, L=block_size, hop=hop, win=analysis_win)
# synthesis
x_r = stft.synthesis(X, L=block_size, hop=hop)
return pra.dB(
np.max(np.abs(x_local[:-block_size + hop, ] -
x_r[block_size - hop:, ])))
def hop_one_sample(D):
if D == 1:
x_local = x[:, 0]
else:
x_local = x[:, :D]
hop = 1
# analysis
analysis_win = pra.hann(block_size)
X = stft.analysis(x_local, L=block_size, hop=hop, win=analysis_win)
# synthesis
synthesis_win = pra.transform.stft.compute_synthesis_window(
analysis_win, hop)
x_r = stft.synthesis(X, L=block_size, hop=hop, win=synthesis_win)
return pra.dB(
np.max(np.abs(x_local[:-block_size + hop, ] -
x_r[block_size - hop:, ])))
def append_one_sample(D):
hop = block_size // 2
n_samples = x.shape[0]
n_frames = n_samples // hop
x_local = x[:n_frames * hop - 1, :]
if D == 1:
x_local = x_local[:, 0]
else:
x_local = x_local[:, :D]
# analysis
analysis_win = pra.hann(block_size)
X = stft.analysis(x_local, L=block_size, hop=hop, win=analysis_win)
# synthesis
x_r = stft.synthesis(X, L=block_size, hop=hop)
return pra.dB(
np.max(
np.abs(x_local[:-block_size + hop, ] -
x_r[block_size - hop:-1, ])))
if args.algo in dereverb_algos:
# for dereverberation algorithms we use the anechoic reference signal
fn_ref = DATA_DIR / rooms[args.room]["anechoic_filenames"][REF_MIC]
else:
fn_ref = DATA_DIR / rooms[args.room]["src_filenames"][REF_MIC]
fs, ref = wavfile.read(fn_ref)
ref = ref.astype(np.float64) / 2**15
# STFT parameters
hop = args.block // 2
win_a = pra.hamming(args.block)
win_s = pra.transform.stft.compute_synthesis_window(win_a, hop)
# STFT
X = stft.analysis(mix, args.block, hop, win=win_a)
t1 = time.perf_counter()
# Separation
if args.algo == "fastmnmf":
Y = algorithms[args.algo](X, n_iter=30)
elif args.algo in dereverb_algos:
if args.p is None:
Y = algorithms[args.algo](
X,
n_iter=15 * args.mics,
n_taps=3,
n_delays=2,
n_components=1,
proj_back=True,
room.add_microphone_array(pra.MicrophoneArray(mic_loc, fs))
room.compute_rir()
#plot spectrograms to check for sweeping echoes
fft_size = 512 # fft size for analysis
fft_hop = 128 # hop between analysis frame
fft_zp = 512 # zero padding
analysis_window = pra.hann(fft_size)
print("Sweeping echo measure for ISM is :")
for n in range(M):
if n == 0:
S = stft.analysis(room.rir[n][0], fft_size, fft_hop, win=analysis_window, zp_back=fft_zp)
f, (ax1, ax2) = plt.subplots(2,1)
ax1.imshow(
pra.dB(S.T),
extent=[0, len(room.rir[n][0]), 0, fs / 2],
vmin=-100,
vmax=0,
origin="lower",
cmap="jet"
)
ax1.set_title("RIR for Mic location " + str(n) + " without random ISM")
ax1.set_ylabel("Frequency")
ax1.set_aspect("auto")
dSNR = pra.dB(room1.direct_snr(mics.center[:, 0], source=0), power=True)
print("The direct SNR for good source is " + str(dSNR))
# remove a bit of signal at the end
n_lim = int(np.ceil(len(input_mic) - t_cut * Fs))
input_clean = signal1[:n_lim]
input_mic = input_mic[:n_lim]
out_DirectMVDR = out_DirectMVDR[:n_lim]
out_RakeMVDR = out_RakeMVDR[:n_lim]
out_DirectPerceptual = out_DirectPerceptual[:n_lim]
out_RakePerceptual = out_RakePerceptual[:n_lim]
# compute time-frequency planes
F0 = stft.analysis(input_clean, fft_size, fft_hop, win=analysis_window, zp_back=fft_zp)
F1 = stft.analysis(input_mic, fft_size, fft_hop, win=analysis_window, zp_back=fft_zp)
F2 = stft.analysis(
out_DirectMVDR, fft_size, fft_hop, win=analysis_window, zp_back=fft_zp
)
F3 = stft.analysis(out_RakeMVDR, fft_size, fft_hop, win=analysis_window, zp_back=fft_zp)
F4 = stft.analysis(
out_DirectPerceptual, fft_size, fft_hop, win=analysis_window, zp_back=fft_zp
)
F5 = stft.analysis(
out_RakePerceptual, fft_size, fft_hop, win=analysis_window, zp_back=fft_zp
)
# (not so) fancy way to set the scale to avoid having the spectrum
# dominated by a few outliers
p_min = 7
def process(args, config):
n_channels, room_id, bss_algo = args
# the name of the algorithm we'll use for bss
bss_algo_name = config["bss_algorithms"][bss_algo]["name"]
if mkl_available:
mkl.set_num_threads_local(1)
ref_mic = config["ref_mic"]
metadata_fn = Path(config["metadata_fn"])
dataset_dir = metadata_fn.parent
with open(config["metadata_fn"], "r") as f:
metadata = json.load(f)
rooms = metadata[f"{n_channels}_channels"]
# the mixtures
fn_mix = dataset_dir / rooms[room_id]["mix_filename"]
fs, mix = load_audio(fn_mix)
# add the noise
sigma_src = np.std(mix)
sigma_n = sigma_src * 10**(-config["snr"] / 20)
mix += np.random.randn(*mix.shape) * sigma_n
# the reference
if bss_algo_name in dereverb_algos:
# for dereverberation algorithms we use the anechoic reference signal
fn_ref = dataset_dir / rooms[room_id]["anechoic_filenames"][
config["ref_mic"]]
else:
fn_ref = dataset_dir / rooms[room_id]["src_filenames"][
config["ref_mic"]]
fs, ref = load_audio(fn_ref)
# STFT parameters
nfft = config["stft"]["nfft"]
hop = config["stft"]["hop"]
if config["stft"]["window"] == "hamming":
win_a = pra.hamming(nfft)
win_s = pra.transform.stft.compute_synthesis_window(win_a, hop)
else:
raise ValueError("Window not implemented")
# STFT
X = stft.analysis(mix, nfft, hop, win=win_a)
# Separation
bss_kwargs = config["bss_algorithms"][bss_algo]["kwargs"]
n_iter_p_ch = config["bss_algorithms"][bss_algo]["n_iter_per_channel"]
runtime_bss = time.perf_counter()
if bss_algo_name == "fastmnmf":
Y = bss_algorithms[bss_algo_name](X,
n_iter=n_iter_p_ch * n_channels,
**bss_kwargs)
elif bss_algo_name in dereverb_algos:
Y, Y_pb, runtime_pb = bss_algorithms[bss_algo_name](
X,
n_iter=n_iter_p_ch * n_channels,
proj_back_both=True,
**bss_kwargs)
# adjust start time to remove the projection back
runtime_bss += runtime_pb
else:
Y = bss_algorithms[bss_algo_name](X,
n_iter=n_iter_p_ch * n_channels,
proj_back=False,
**bss_kwargs)
runtime_bss = time.perf_counter() - runtime_bss
results = [{
"bss_runtime": {
"bss_algo": bss_algo,
"runtime": runtime_bss,
}
}]
t = {
"room_id": room_id,
"n_channels": n_channels,
"bss_algo": bss_algo,
"proj_algo": None,
"runtime": 0.0,
"sdr": None,
"sir": None,
"p": None,
"q": None,
"n_iter": 1,
}
# Evaluation of raw signal
t["proj_algo"] = "None"
y, sdr, sir, _ = reconstruct_evaluate(ref,
Y,
nfft,
hop,
win=win_s,
si_metric=config["si_metric"])
t["sdr"], t["sir"] = sdr.tolist(), sir.tolist()
results.append(t.copy())
# projection back
t["proj_algo"] = "projection_back"
if bss_algo in dereverb_algos:
Z = Y_pb
else:
runtime_pb = time.perf_counter()
Z = bss_scale.projection_back(Y, X[:, :, ref_mic])
runtime_pb = time.perf_counter() - runtime_pb
y, sdr, sir, _ = reconstruct_evaluate(ref,
Z,
nfft,
hop,
win=win_s,
si_metric=config["si_metric"])
t["sdr"], t["sir"] = sdr.tolist(), sir.tolist()
t["runtime"] = runtime_pb
results.append(t.copy())
# minimum distortion
lo, hi, n_steps = config["minimum_distortion"]["p_list"]
p_vals = np.linspace(lo, hi, n_steps)
kwargs = config["minimum_distortion"]["kwargs"]
for ip, p in enumerate(p_vals):
for q in p_vals[ip:]:
t["p"], t["q"], t["proj_algo"] = (
f"{p:.1f}",
f"{q:.1f}",
"minimum_distortion",
)
runtime_md = time.perf_counter()
Z, t["n_iter"] = bss_scale.minimum_distortion(Y,
X[:, :, ref_mic],
p=p,
q=q,
**kwargs)
runtime_md = time.perf_counter() - runtime_md
y, sdr, sir, _ = reconstruct_evaluate(
ref, Z, nfft, hop, win=win_s, si_metric=config["si_metric"])
t["sdr"] = sdr.tolist()
t["sir"] = sir.tolist()
t["runtime"] = runtime_md
results.append(t.copy())
return results