def _test(method='AuxLaplaceIVA'):
np.random.seed(111)
# Room impulse response
sr = 16000
reverb = 0.16
duration = 0.5
samples = int(duration * sr)
mic_intervals = [8, 8, 8, 8, 8, 8, 8]
mic_indices = [2, 5]
degrees = [60, 300]
titles = ['man-16000', 'woman-16000']
mixed_signal = _convolve_mird(titles, reverb=reverb, degrees=degrees, mic_intervals=mic_intervals, mic_indices=mic_indices, samples=samples)
n_channels, T = mixed_signal.shape
# STFT
fft_size, hop_size = 2048, 1024
mixture = stft(mixed_signal, fft_size=fft_size, hop_size=hop_size)
# IVA
lr = 0.1
n_sources = len(titles)
iteration = 200
if method == 'GradLaplaceIVA':
iva = GradLaplaceIVA(lr=lr)
iteration = 5000
elif method == 'NaturalGradLaplaceIVA':
iva = NaturalGradLaplaceIVA(lr=lr)
iteration = 200
elif method == 'AuxLaplaceIVA':
iva = AuxLaplaceIVA()
iteration = 50
else:
raise ValueError("Not support method {}".format(method))
estimation = iva(mixture, iteration=iteration)
estimated_signal = istft(estimation, fft_size=fft_size, hop_size=hop_size, length=T)
print("Mixture: {}, Estimation: {}".format(mixed_signal.shape, estimated_signal.shape))
for idx in range(n_sources):
_estimated_signal = estimated_signal[idx]
write_wav("data/IVA/{}/mixture-{}_estimated-iter{}-{}.wav".format(method, sr, iteration, idx), signal=_estimated_signal, sr=sr)
plt.figure()
plt.plot(iva.loss, color='black')
plt.xlabel('Iteration')
plt.ylabel('Loss')
plt.savefig('data/IVA/{}/loss.png'.format(method), bbox_inches='tight')
plt.close()
def _test(method, n_bases=10, partitioning=False):
np.random.seed(111)
# Room impulse response
sr = 16000
reverb = 0.16
duration = 0.5
samples = int(duration * sr)
mic_intervals = [8, 8, 8, 8, 8, 8, 8]
mic_indices = [2, 5]
degrees = [60, 300]
titles = ['man-16000', 'woman-16000']
mixed_signal = _convolve_mird(titles, reverb=reverb, degrees=degrees, mic_intervals=mic_intervals, mic_indices=mic_indices, samples=samples)
n_sources, T = mixed_signal.shape
# STFT
fft_size, hop_size = 2048, 1024
mixture = stft(mixed_signal, fft_size=fft_size, hop_size=hop_size)
# ILRMA
n_channels = len(titles)
iteration = 200
if method == 'Gauss':
ilrma = GaussILRMA(n_bases=n_bases, partitioning=partitioning)
elif method == 't':
ilrma = tILRMA(n_bases=n_bases, partitioning=partitioning)
else:
raise ValueError("Not support {}-ILRMA.".format(method))
estimation = ilrma(mixture, iteration=iteration)
estimated_signal = istft(estimation, fft_size=fft_size, hop_size=hop_size, length=T)
print("Mixture: {}, Estimation: {}".format(mixed_signal.shape, estimated_signal.shape))
for idx in range(n_channels):
_estimated_signal = estimated_signal[idx]
write_wav("data/ILRMA/{}ILMRA/partitioning{}/mixture-{}_estimated-iter{}-{}.wav".format(method, int(partitioning), sr, iteration, idx), signal=_estimated_signal, sr=sr)
plt.figure()
plt.plot(ilrma.loss, color='black')
plt.xlabel('Iteration')
plt.ylabel('Loss')
plt.savefig('data/ILRMA/{}ILMRA/partitioning{}/loss.png'.format(method, int(partitioning)), bbox_inches='tight')
plt.close()
def _test(method='DSBF'):
# Room impulse response
sr = 16000
reverb = 0.16
duration = 0.5
samples = int(duration * sr)
mic_intervals = [3, 3, 3, 8, 3, 3, 3]
mic_indices = [0, 1, 2, 3, 4, 5, 6, 7]
mic_position = np.array([[0.13, 0], [0.10, 0], [0.07, 0], [0.04, 0],
[-0.04, 0], [-0.07, 0], [-0.10, 0], [-0.13, 0]])
degrees = [0, 90]
titles = ['man-16000', 'woman-16000']
n_sources, n_channels = len(degrees), len(mic_indices)
mixed_signal = _convolve_mird(titles,
reverb=reverb,
degrees=degrees,
mic_intervals=mic_intervals,
mic_indices=mic_indices,
samples=samples)
_, T = mixed_signal.shape
# STFT
fft_size, hop_size = 2048, 1024
n_bins = fft_size // 2 + 1
frequency = np.arange(0, n_bins) * sr / fft_size
mixture = stft(mixed_signal, fft_size=fft_size,
hop_size=hop_size) # (n_channels, n_bins, n_frames)
# Steeing vectors
degrees = np.array(degrees) / 180 * np.pi
x_source, y_source = np.sin(degrees), np.cos(degrees) # (n_sources,)
source_position = np.vstack([x_source,
y_source]).transpose(1, 0) # (n_sources, 2)
steering_vector = np.exp(
2j * np.pi * frequency[:, np.newaxis, np.newaxis] *
np.sum(source_position * mic_position[:, np.newaxis, :], axis=2) /
sound_speed) # (n_bins, n_channels, n_sources)
steering_vector = steering_vector / np.sqrt(len(mic_indices))
if method == 'DSBF':
beamformer = DelaySumBeamformer(steering_vector=steering_vector)
elif method == 'MVDR':
beamformer = MVDRBeamformer(steering_vector=steering_vector)
else:
raise NotImplementedError("Not support {} beamformer".format(method))
estimation = beamformer(mixture)
spectrogram = np.abs(estimation)
log_spectrogram = 10 * np.log10(spectrogram**2)
N, F_bin, T_bin = log_spectrogram.shape
t = np.arange(T_bin + 1)
f = np.arange(F_bin + 1)
for n in range(N):
plt.figure()
plt.pcolormesh(t, f, log_spectrogram[n], cmap='jet')
plt.savefig("data/Beamform/{}/specrtogram-{}.png".format(method, n),
bbox_inches='tight')
plt.close()
estimated_signal = istft(estimation,
fft_size=fft_size,
hop_size=hop_size,
length=T)
print("Mixture: {}, Estimation: {}".format(mixed_signal.shape,
estimated_signal.shape))
for idx in range(n_sources):
_estimated_signal = estimated_signal[idx]
write_wav("data/Beamform/{}/mixture-{}_estimated-{}.wav".format(
method, sr, idx),
signal=_estimated_signal,
sr=sr)
def _test(metric='EUC'):
np.random.seed(111)
fft_size, hop_size = 1024, 256
n_bases = 6
iteration = 100
signal, sr = read_wav("data/single-channel/music-8000.wav")
T = len(signal)
spectrogram = stft(signal, fft_size=fft_size, hop_size=hop_size)
amplitude = np.abs(spectrogram)
power = amplitude**2
if metric == 'EUC':
nmf = EUCNMF(n_bases)
elif metric == 'IS':
nmf = ISNMF(n_bases)
elif metric == 'KL':
nmf = KLNMF(n_bases)
else:
raise NotImplementedError("Not support {}-NMF".format(metric))
nmf.update(power, iteration=iteration)
amplitude[amplitude < EPS] = EPS
estimated_power = nmf.base @ nmf.activation
estimated_amplitude = np.sqrt(estimated_power)
ratio = estimated_amplitude / amplitude
estimated_spectrogram = ratio * spectrogram
estimated_signal = istft(estimated_spectrogram,
fft_size=fft_size,
hop_size=hop_size,
length=T)
estimated_signal = estimated_signal / np.abs(estimated_signal).max()
write_wav("data/NMF/{}/music-8000-estimated-iter{}.wav".format(
metric, iteration),
signal=estimated_signal,
sr=8000)
power[power < EPS] = EPS
log_spectrogram = 10 * np.log10(power)
plt.figure()
plt.pcolormesh(log_spectrogram, cmap='jet')
plt.colorbar()
plt.savefig('data/NMF/spectrogram.png', bbox_inches='tight')
plt.close()
for idx in range(n_bases):
estimated_power = nmf.base[:, idx:idx + 1] @ nmf.activation[idx:idx +
1, :]
estimated_amplitude = np.sqrt(estimated_power)
ratio = estimated_amplitude / amplitude
estimated_spectrogram = ratio * spectrogram
estimated_signal = istft(estimated_spectrogram,
fft_size=fft_size,
hop_size=hop_size,
length=T)
estimated_signal = estimated_signal / np.abs(estimated_signal).max()
write_wav("data/NMF/{}/music-8000-estimated-iter{}-base{}.wav".format(
metric, iteration, idx),
signal=estimated_signal,
sr=8000)
estimated_power[estimated_power < EPS] = EPS
log_spectrogram = 10 * np.log10(estimated_power)
plt.figure()
plt.pcolormesh(log_spectrogram, cmap='jet')
plt.colorbar()
plt.savefig(
'data/NMF/{}/estimated-spectrogram-iter{}-base{}.png'.format(
metric, iteration, idx),
bbox_inches='tight')
plt.close()
plt.figure()
plt.plot(nmf.loss, color='black')
plt.xlabel('Iteration')
plt.ylabel('Loss')
plt.savefig('data/NMF/{}/loss.png'.format(metric), bbox_inches='tight')
plt.close()