def split_frame_m():
c = 343
frame_count = 2048
data, fs = load_audio_data(
r'D:\projects\pyprojects\soundphase\calib\0\0.wav', 'wav')
skip_time = int(fs * 1)
data = data[skip_time:, :-1].T
# search unit circle
level = 4
grid: np.ndarray = np.load(rf'grid/{level}.npz')['grid']
# mic mem pos
pos = cons_uca(0.043)
for i in range(0, data.shape[1], frame_count):
data_seg = data[:, i:i + frame_count]
# 噪声不做,随便写的
if np.max(abs(fft(data_seg[0] / len(data_seg[0])))) < 1:
continue
print('time: ', (skip_time + i) / fs)
t1 = time.time()
pair_tau, stack_fft = calculate_stack_fft_and_pairs_tau(
data_seg, pos, grid, c)
t2 = time.time()
print('previous calculation time consumption: ', t2 - t1)
t1 = time.time()
E = srp_phat_m(data_seg, pos, stack_fft, pair_tau, fs)
# E = srp_phat_previous_tau(data_seg, pos, grid, pairs_tau, fs)
t2 = time.time()
print('srp_phat time consumption: ', t2 - t1)
sdevc = grid[np.argmax(E, axis=1)] # source direction vector
# print(sdevc)
print('angle of max val: ', np.rad2deg(vec2theta(sdevc)))
print('=' * 50)
def split_frame():
c = 343
frame_count = 1024
data, fs = load_audio_data(
r'D:\projects\pyprojects\gesturerecord\location\sound\0.wav', 'wav')
skip_time = int(fs * 1)
data = data[skip_time:, :-1].T
# search unit circle
level = 3
grid: np.ndarray = np.load(rf'grid/{level}_north.npz')['grid']
# mic mem pos
pos = cons_uca(0.043)
# calculate tau previously
# pairs_tau = calculate_pairs_tau(pos, grid, c)
for i in range(0, data.shape[1], frame_count):
data_seg = data[:, i:i + frame_count]
# 噪声不做,随便写的
if np.max(abs(fft(data_seg[0] / len(data_seg[0])))) < 10:
continue
print('time: ', (skip_time + i) / fs)
t1 = time.time()
E = srp_phat(data_seg, pos, grid, c, fs)
# E = srp_phat_muti_thread(data, pos, grid, c, fs)
# E = srp_phat_previous_tau(data_seg, pos, grid, pairs_tau, fs)
t2 = time.time()
print('srp_phat time consumption: ', t2 - t1)
sdevc = grid[np.argmax(E, axis=1)] # source direction vector
# print(sdevc)
print('angle of max val: ', np.rad2deg(vec2theta(sdevc)))
print('=' * 50)
def srp_fft_denoise_test():
frame_len = 2048
# 加窗窗口
window = np.hanning(frame_len)
# window = np.ones(frame_len)
# window = np.hamming(frame_len)
data, fs = load_audio_data(r'D:\projects\pyprojects\gesturerecord\location\20khz\0.wav', 'wav')
data = data[fs * 1:, :-1].T
data_filter = butter_bandpass_filter(data, 19e3, 21e3)
# for f in data_filter:
# normalized_signal_fft(f)
t_fs_n = 40000
noise_data = data_filter[:, t_fs_n:t_fs_n+frame_len]
t_fs_t = 48000 * 5 + 1000
test_data = data_filter[:, t_fs_t:t_fs_t+frame_len]
# denoise_fft(fft(test_data)/frame_len, fft(noise_data)/frame_len)
mic_array_pos = cons_uca(0.043)
c = 343
level = 4
grid = np.load(rf'grid/{level}_north.npz')['grid']
noise_fft = fft(noise_data * window)
E = srp_phat_denoise(test_data, mic_array_pos, grid, c, fs, noise_fft, window)
sdevc = grid[np.argmax(E, axis=1)] # source direction vector
# print(sdevc)
print('angle of max val: ', np.rad2deg(vec2theta(sdevc)))
print('=' * 50)
def extract_phasedata_from_audio_special_for_onemic(audio_file,
phasedata_save_file,
audio_type='wav',
mic_array=True):
origin_data, fs = load_audio_data(audio_file, audio_type)
fs = fs # 采样率
data = origin_data.reshape((-1, 8))
data = data.T # shape = (num_of_channels, all_frames)
data = data[:, int(fs * DELAY_TIME):]
mic_num = 0
# 只用一个mic
data = data[mic_num, :]
data = data.reshape((1, -1))
# 开始处理数据
t = 0
magnti_list = []
for i in range(NUM_OF_FREQ):
fc = F0 + i * STEP
data_filter = butter_bandpass_filter(data, fc - 150, fc + 150)
I_raw, Q_raw = get_cos_IQ_raw(data_filter, fc, fs)
# 滤波+下采样
I = move_average_overlap_filter(I_raw)
Q = move_average_overlap_filter(Q_raw)
# denoise
decompositionQ = seasonal_decompose(Q.T, period=10, two_sided=False)
trendQ = decompositionQ.trend
decompositionI = seasonal_decompose(I.T, period=10, two_sided=False)
trendI = decompositionI.trend
trendQ = trendQ.T
trendI = trendI.T
assert trendI.shape == trendQ.shape
if len(trendI.shape) == 1:
trendI = trendI.reshape((1, -1))
trendQ = trendQ.reshape((1, -1))
trendQ = trendQ[:, 10:]
trendI = trendI[:, 10:]
magnti = get_phase(trendI, trendQ) # 这里的展开目前没什么效果
# plt.plot(magnti[0])
# plt.show()
assert magnti.shape[1] > 1
# 用diff，和两次diff
magnti_list.append(np.diff(magnti)[:, :-1])
# plt.plot(np.diff(magnti).reshape(-1))
# plt.show()
magnti_list.append(np.diff(np.diff(magnti)))
merged_u_p = np.array(magnti_list).reshape((NUM_OF_FREQ * 1 * 2, -1))
print(merged_u_p.shape)
# 压缩便于保存
flattened_m_u_p = merged_u_p.flatten()
# 由于长短不一，不能放在一起
# np.savetxt(dataset_save_file, flattened_m_u_p.reshape(1, -1))
np.savez_compressed(phasedata_save_file, phasedata=flattened_m_u_p)
return 1
def compare():
LENG = 2048
data, fs = load_audio_data(r'D:\projects\pyprojects\gesturerecord\location\sound\0.wav', 'wav')
# data = butter_bandpass_filter(data.T, 15e3, 23e3)
data = data.T
t = 3
data = data[:-1, int(fs * t):int(fs * t) + LENG]
#
# for i, d in enumerate(data):
# plt.subplot(4, 2, i + 1)
# plt.plot(d)
# plt.show()
i = 0
j = 1
a = data[i]
b = data[j]
y = gcc_phat(a, b)
print('ifft max ccor val: ', np.max(y))
print('ifft delay of sample num: ', np.argmax(y))
plt.figure()
plt.plot(y)
plt.title('ifft gcc')
mic_array_pos = cons_uca(0.043)
c = 343
level = 4
grid = np.load(rf'grid/{level}.npz')['grid']
tau = get_steering_vector(mic_array_pos[i], mic_array_pos[j], c, grid)
R = gcc_phat_search(a, b, fs, tau)
# 画射线图
# plot_angspect(R[0], grid, percentile=99)
# r = R[0]
# sorted_arg = np.argsort(r)[::-1]
# print(r[sorted_arg])
print('gccphat search max val: ', np.max(R))
print('gccphat delay of sample num: ', tau[np.argmax(R)] * fs)
max_p = grid[np.argmax(R)]
print('point of max val: ', max_p)
print('angle of max val: ', np.rad2deg(vec2theta([max_p])))
plt.figure()
plt.plot(R.reshape(-1))
plt.title('gcc search')
# plt.figure()
# plt.plot(np.correlate(a,b,'full'))
plt.show()
def generate_training_data_pcm(audio_file, dataset_save_file):
origin_data, fs = load_audio_data(audio_file, 'pcm')
nchannels = 1 # 声道数
fs = fs # 采样率
# 开始处理数据
t = 0
f0 = 17350
for win in range(CHUNK, len(origin_data), CHUNK):
# 读取下一段数据
data = origin_data[win - CHUNK:win + CHUNK]
t = t + CHUNK / fs
# print(f"time:{t}s")
# 由于麦克风的原因只看2s之后的
if t < DELAY_TIME:
continue
unwrapped_phase_list = []
data = data.reshape((-1, nchannels))
data = data.T # shape = (num_of_channels, 2 * CHUNK)
if data.shape[1] < 2 * CHUNK:
continue
# 处理数据，这里可以优化，还需要验证其正确性
for i in range(NUM_OF_FREQ):
fc = f0 + i * STEP
data_filter = butter_bandpass_filter(data, fc - 250, fc + 250)
I, Q = get_cos_IQ(data_filter, fc, fs)
unwrapped_phase = get_phase(I, Q) # 这里的展开目前没什么效果
# plt.plot(unwrapped_phase[0])
# plt.show()
# 通过标准差判断是否在运动
assert unwrapped_phase.shape[1] > 1
u_p_stds = np.std(unwrapped_phase, axis=1)
print(fc, np.mean(u_p_stds))
if np.mean(u_p_stds) > STD_THRESHOLD:
unwrapped_phase_list.append(unwrapped_phase)
# 把8个频率的合并成一个矩阵 shape = (num_of_channels * NUM_OF_FREQ, CHUNK)
if len(unwrapped_phase_list) != NUM_OF_FREQ:
continue
print(f"time:{t}s")
merged_u_p = np.vstack(([u_p for u_p in unwrapped_phase_list]))
# 压缩便于保存
flattened_m_u_p = merged_u_p.flatten()
with open(dataset_save_file, 'ab') as f:
np.savetxt(f, flattened_m_u_p.reshape(1, -1))
def fft_denoise_test():
frame_len = 2048
# 加窗窗口
window = np.hanning(frame_len)
# window = np.ones(frame_len)
# window = np.hamming(frame_len)
data, fs = load_audio_data(r'D:\projects\pyprojects\gesturerecord\location\20khz\0.wav', 'wav')
data = data[fs * 1:, :-1].T
data_filter = butter_bandpass_filter(data, 19e3, 21e3)
# for f in data_filter:
# normalized_signal_fft(f)
t_fs_n = 40000
noise_data = data_filter[:, t_fs_n:t_fs_n+frame_len]
t_fs_t = 48000 * 5 + 1000
test_data = data_filter[:, t_fs_t:t_fs_t+frame_len]
# denoise_fft(fft(test_data)/frame_len, fft(noise_data)/frame_len)
i = 0
j = 1
mic_array_pos = cons_uca(0.043)
c = 343
level = 4
grid = np.load(rf'grid/{level}.npz')['grid']
tau = get_steering_vector(mic_array_pos[i], mic_array_pos[j], c, grid)
noise_fft = fft(noise_data * window)
R = gcc_phat_search_fft_denoise(test_data[i], test_data[j], fs, tau, noise_fft[i], noise_fft[j], window)
print('gccphat search max val: ', np.max(R))
print('gccphat delay of sample num: ', tau[np.argmax(R)] * fs)
max_p = grid[np.argmax(R)]
print('point of max val: ', max_p)
print('angle of max val: ', np.rad2deg(vec2theta([max_p])))
plt.figure()
plt.plot(R.reshape(-1))
plt.title('gcc search')
# plt.figure()
# plt.plot(np.correlate(a,b,'full'))
plt.show()
def beamform_on_raw_audio_data(filename):
data, fs = load_audio_data(filename, 'wav')
data = data.T
data = data[:7, int(fs * DELAY_TIME):]
a_angel = 240
e_angel = 0
two_d_angel = [[np.deg2rad(a_angel), np.deg2rad(e_angel)]]
c = 343
spacing = 0.043
mic_array_pos = cons_uca(spacing)
sd = steering_plane_wave(mic_array_pos, c, two_d_angel)
beamformed_data = beamform_real(data, sd).reshape(1, -1)
beamformed_data_2 = ump_8_beamform(data, 48000, two_d_angel).reshape(1, -1)
phase_list = []
for i in range(NUM_OF_FREQ):
fc = F0 + i * STEP
data_filter = butter_bandpass_filter(data, fc - 150, fc + 150)
I_raw, Q_raw = get_cos_IQ_raw(data_filter, fc, fs)
# 滤波+下采样
I = move_average_overlap_filter(I_raw)
Q = move_average_overlap_filter(Q_raw)
decompositionQ = seasonal_decompose(Q.T, period=10, two_sided=False)
trendQ = decompositionQ.trend
decompositionI = seasonal_decompose(I.T, period=10, two_sided=False)
trendI = decompositionI.trend
trendQ = trendQ.T
trendI = trendI.T
# trendQ = Q
# trendI = I
assert trendI.shape == trendQ.shape
if len(trendI.shape) == 1:
trendI = trendI.reshape((1, -1))
trendQ = trendQ.reshape((1, -1))
# trendQ = trendQ[:, 10:]
# trendI = trendI[:, 10:]
trendQ = Q
trendI = I
# draw_circle(trendI[0], trendQ[0])
raw_phase = get_phase(trendI, trendQ)
'''
对相位beamform
'''
noise = np.mean(raw_phase[:, 20:60], axis=1).reshape(-1, 1)
print(noise.shape)
raw_phase_denoised = raw_phase - noise
bphase_denoised = beamform_real(raw_phase_denoised, sd).reshape(1, -1)
def normalize_max_min(x):
max = np.max(x)
min = np.min(x)
return (x - min) / (max - min)
'''
对beamformed signal求相位
'''
beamformed_data_filter = butter_bandpass_filter(
beamformed_data, fc - 150, fc + 150)
bI_raw, bQ_raw = get_cos_IQ_raw(beamformed_data_filter, fc, fs)
bI = move_average_overlap_filter(bI_raw)
bQ = move_average_overlap_filter(bQ_raw)
bphase = get_phase(bI, bQ)
beamformed_data_filter = butter_bandpass_filter(
beamformed_data_2, fc - 150, fc + 150)
bI_raw, bQ_raw = get_cos_IQ_raw(beamformed_data_filter, fc, fs)
bI = move_average_overlap_filter(bI_raw)
bQ = move_average_overlap_filter(bQ_raw)
bphase_2 = get_phase(bI, bQ)
plt.figure()
plt.subplot(2, 1, 1)
plt.plot(raw_phase[0])
plt.subplot(2, 1, 2)
plt.plot(bphase_denoised[0])
plt.show()
def beamform_after_IQ(filename, start, dur):
data, fs = load_audio_data(filename, 'wav')
data = data.T
data = data[:7, int(fs * DELAY_TIME):]
# data = data[:7, start:start+dur]
phase_list = []
for i in range(NUM_OF_FREQ):
fc = F0 + i * STEP
data_filter = butter_bandpass_filter(data, fc - 150, fc + 150)
I_raw, Q_raw = get_cos_IQ_raw(data_filter, fc, fs)
# 滤波+下采样
I = move_average_overlap_filter(I_raw)
Q = move_average_overlap_filter(Q_raw)
# I = butter_lowpass_filter(I_raw, 150)
# Q = butter_lowpass_filter(Q_raw, 150)
# I = I[:, 5:-5]
# Q = Q[:, 5:-5]
# plt.plot(I[0][5:-5])
# plt.plot(Q[0][5:-5])
# plt.show()
# denoise
decompositionQ = seasonal_decompose(Q.T, period=10, two_sided=False)
trendQ = decompositionQ.trend
decompositionI = seasonal_decompose(I.T, period=10, two_sided=False)
trendI = decompositionI.trend
trendQ = trendQ.T
trendI = trendI.T
# trendQ = Q
# trendI = I
assert trendI.shape == trendQ.shape
if len(trendI.shape) == 1:
trendI = trendI.reshape((1, -1))
trendQ = trendQ.reshape((1, -1))
# trendQ = trendQ[:, 10:]
# trendI = trendI[:, 10:]
trendQ = Q
trendI = I
# draw_circle(trendI[0], trendQ[0])
raw_phase = get_phase(trendI, trendQ)
exp_phase = trendI + 1j * trendQ
'''
去除噪声，减去平均值（尝试）
'''
mean_noise = np.mean(raw_phase[:1000], axis=1)
# beamform
a_angel = 120
e_angel = 0
two_d_angel = [[np.deg2rad(a_angel), np.deg2rad(e_angel)]]
c = 343
spacing = 0.043
mic_array_pos = cons_uca(spacing)
# sp = music(data, mic_array_pos, fc, c, np.arange(0, 360), np.arange(0, 30), 1)
# # plt.plot(sp.reshape(-1))
# plt.pcolormesh(sp)
# plt.show()
azumi = (0, 360)
eleva = (0, 90)
# beamscan_spectrum = np.zeros((azumi[1] - azumi[0], eleva[1] - eleva[0]))
# # 估计
# for angel_1 in range(azumi[0], azumi[1]):
# for angel_2 in range(eleva[0], eleva[1]):
# two_angel = [[np.deg2rad(angel_1), np.deg2rad(angel_2)]]
# sd = steering_plane_wave(mic_array_pos, c, two_angel)
# adjust = np.exp(-1j * 2 * np.pi * fc * sd)
# syn_signals = exp_phase * adjust.T
# # syn_signals = np.real(syn_signals)
# beamformed_signal = np.sum(syn_signals, axis=0)
# beamscan_spectrum[angel_1][angel_2] = np.sum(abs(beamformed_signal))
# # return beamscan_spectrum
# # plt.pcolormesh(beamscan_spectrum)
# # plt.show()
# plt.plot(beamscan_spectrum[:, 0])
# plt.grid()
# plt.show()
sd = steering_plane_wave(mic_array_pos, c, two_d_angel)
adjust = np.exp(-1j * 2 * np.pi * fc * sd).T
assert exp_phase.shape[0] == adjust.shape[0]
syn_signals = exp_phase * adjust
beamformed_signal = np.sum(syn_signals, axis=0).reshape(1, -1)
phase = get_phase(np.real(beamformed_signal),
np.imag(beamformed_signal))
plt.show()
plt.figure()
plt.subplot(2, 1, 1)
plt.plot(raw_phase[0])
plt.subplot(2, 1, 2)
plt.plot(phase[0])
plt.show()
phase_list.append(phase)
def extract_magndata_from_beamformed_audio(audio_file,
phasedata_save_file,
audio_type='pcm',
mic_array=False):
origin_data, fs = load_audio_data(audio_file, audio_type)
fs = fs # 采样率
# 已经reshape过了为什么还要reshape
if mic_array:
data = origin_data.reshape((-1, N_CHANNELS + 1))
else:
data = origin_data.reshape((-1, N_CHANNELS))
data = data.T # shape = (num_of_channels, all_frames)
data = data[:, int(fs * DELAY_TIME):]
if mic_array:
# 第八个声道不要
data = data[:7, :]
# beamform,角度？
data = ump_8_beamform(data, fs, angel=[[np.pi * 4 / 3, 0]])
assert data.shape[0] == 1
# 开始处理数据
t = 0
magnti_list = []
for i in range(NUM_OF_FREQ):
fc = F0 + i * STEP
data_filter = butter_bandpass_filter(data, fc - 150, fc + 150)
I_raw, Q_raw = get_cos_IQ_raw(data_filter, fc, fs)
# 滤波+下采样
I = move_average_overlap_filter(I_raw)
Q = move_average_overlap_filter(Q_raw)
# denoise
decompositionQ = seasonal_decompose(Q.T, period=10, two_sided=False)
trendQ = decompositionQ.trend
decompositionI = seasonal_decompose(I.T, period=10, two_sided=False)
trendI = decompositionI.trend
trendQ = trendQ.T
trendI = trendI.T
assert trendI.shape == trendQ.shape
if len(trendI.shape) == 1:
trendI = trendI.reshape((1, -1))
trendQ = trendQ.reshape((1, -1))
trendQ = trendQ[:, 10:]
trendI = trendI[:, 10:]
magnti = get_magnitude(trendI, trendQ) # 这里的展开目前没什么效果
# plt.plot(np.diff(magnti[0].reshape(-1)))
# plt.show()
assert magnti.shape[1] > 1
# 用diff，和两次diff
magnti_list.append(np.diff(magnti)[:, :-1])
# plt.plot(np.diff(magnti).reshape(-1))
# plt.show()
magnti_list.append(np.diff(np.diff(magnti)))
merged_u_p = np.array(magnti_list).reshape((NUM_OF_FREQ * 1 * 2, -1))
print(merged_u_p.shape)
# 压缩便于保存
flattened_m_u_p = merged_u_p.flatten()
# 由于长短不一，不能放在一起
# np.savetxt(dataset_save_file, flattened_m_u_p.reshape(1, -1))
np.savez_compressed(phasedata_save_file, phasedata=flattened_m_u_p)
return 1
def extract_phasedata_from_audio(audio_file,
phasedata_save_file,
audio_type='pcm',
mic_array=False):
'''
目前使用的是对相位取一次差分
:param audio_file:
:param phasedata_save_file:
:param audio_type:
:param mic_array:
:return:
'''
origin_data, fs = load_audio_data(audio_file, audio_type)
fs = fs # 采样率
# data = origin_data[int(fs * DELAY_TIME):]
# data = data.reshape((-1, N_CHANNELS))
# data = data.T # shape = (num_of_channels, all_frames)
if mic_array:
data = origin_data.reshape((-1, N_CHANNELS + 1))
else:
data = origin_data.reshape((-1, N_CHANNELS))
data = data.T # shape = (num_of_channels, all_frames)
data = data[:, int(fs * DELAY_TIME):]
if mic_array:
# 第八个声道不要
data = data[:7, :]
# 开始处理数据
t = 0
unwrapped_phase_list = []
for i in range(NUM_OF_FREQ):
fc = F0 + i * STEP
data_filter = butter_bandpass_filter(data, fc - 150, fc + 150)
I_raw, Q_raw = get_cos_IQ_raw(data_filter, fc, fs)
# 滤波+下采样
I = move_average_overlap_filter(I_raw)
Q = move_average_overlap_filter(Q_raw)
# denoise
decompositionQ = seasonal_decompose(Q.T, period=10, two_sided=False)
trendQ = decompositionQ.trend
decompositionI = seasonal_decompose(I.T, period=10, two_sided=False)
trendI = decompositionI.trend
trendQ = trendQ.T
trendI = trendI.T
assert trendI.shape == trendQ.shape
if len(trendI.shape) == 1:
trendI = trendI.reshape((1, -1))
trendQ = trendQ.reshape((1, -1))
trendQ = trendQ[:, 10:]
trendI = trendI[:, 10:]
unwrapped_phase = get_phase(trendI, trendQ) # 这里的展开目前没什么效果
# plt.plot(unwrapped_phase[0])
# plt.show()
assert unwrapped_phase.shape[1] > 1
# 用diff，和两次diff
unwrapped_phase_list.append(np.diff(unwrapped_phase)[:, :-1])
# plt.plot(np.diff(unwrapped_phase).reshape(-1))
# plt.show()
# unwrapped_phase_list.append(np.diff(np.diff(unwrapped_phase)))
merged_u_p = np.array(unwrapped_phase_list).reshape(
(NUM_OF_FREQ * N_CHANNELS * 1, -1))
print(merged_u_p.shape)
# 压缩便于保存
flattened_m_u_p = merged_u_p.flatten()
# 由于长短不一，不能放在一起
# np.savetxt(dataset_save_file, flattened_m_u_p.reshape(1, -1))
np.savez_compressed(phasedata_save_file, phasedata=flattened_m_u_p)
return N_CHANNELS
