class VirtualStereoMic(object):
def __init__(self, radius=0.1):
self.radius = radius
self.lmic_pos = Position(0, radius, 0, 'cartesian')
self.rmic_pos = Position(0, -radius, 0, 'cartesian')
def binauralize(self, sources):
if isinstance(sources, PositionalSource):
sources = [sources]
l_signal, r_signal = 0, 0.
for src in sources:
l_dist = np.sqrt(((src.position.coords('cartesian') - self.lmic_pos.coords('cartesian'))**2).sum())
r_dist = np.sqrt(((src.position.coords('cartesian') - self.rmic_pos.coords('cartesian'))**2).sum())
# Time delay
l_delay, r_delay = int(l_dist / C * src.sample_rate), int(r_dist / C * src.sample_rate)
# Attenuation is frequency dependent, but lets simplify.
l_attn, r_attn = 1 / (1. + l_dist), 1 / (1. + r_dist)
l_signal += l_attn * shift(src.signal, l_delay, cval=0.) / len(sources)
r_signal += r_attn * shift(src.signal, r_delay, cval=0.) / len(sources)
return np.stack((l_signal, r_signal), axis=1)
def binauralize_frame(self, sources, output, frame_no):
if isinstance(sources, PositionalSource):
sources = [sources]
for src in sources:
l_dist = np.sqrt(((src.position.coords('cartesian') - self.lmic_pos.coords('cartesian'))**2).sum())
r_dist = np.sqrt(((src.position.coords('cartesian') - self.rmic_pos.coords('cartesian'))**2).sum())
# Time delay
l_delay, r_delay = int(l_dist / C * src.sample_rate), int(r_dist / C * src.sample_rate)
# Attenuation is frequency dependent, but lets simplify.
l_attn, r_attn = 1 / (1. + l_dist), 1 / (1. + r_dist)
if frame_no-l_delay >= 0:
output[frame_no, 0] += l_attn * src.signal[frame_no-l_delay] / len(sources)
if frame_no-r_delay >= 0:
output[frame_no, 1] += r_attn * src.signal[frame_no-r_delay] / len(sources)
def audio_crop_freq_sep(ambix, center, sigma=9, thr=0.7):
import librosa
from scipy import ndimage
center_stft = librosa.core.stft(project_audio(ambix, center=center)[0])
center_stft_smooth = ndimage.gaussian_filter(np.abs(center_stft), sigma=sigma)
other_stft_smooth = []
for phi in np.linspace(-pi, pi, 16):
for nu in np.linspace(-pi / 2, pi / 2, 8):
p = Position(phi, nu, 1., 'polar')
if (p.coords('cartesian') * center.coords('cartesian')).sum() > math.cos(pi / 2):
continue
stft = librosa.core.stft(project_audio(ambix, center=p)[0])
other_stft_smooth += [ndimage.gaussian_filter(np.abs(stft), sigma=sigma)]
other_stft_smooth = np.stack(other_stft_smooth, 0)
rank = (np.abs(center_stft_smooth[np.newaxis]) > np.abs(other_stft_smooth)).sum(0) / other_stft_smooth.shape[0]
stft = librosa.core.stft(project_audio(ambix, center=center)[0])
stft[rank < thr] = 0
wav = librosa.core.istft(stft)
return wav