def lpc_to_mag(v_lpc, fft_len=4096):
'''
Computed the magnitude spectrum from LPC coefficients using approximation by FFT method.
'''
v_imp = np.r_[1, np.zeros(fft_len-1)]
v_imp_filt = lfilter(np.array([1.0]), v_lpc, v_imp)
v_mag = np.absolute(np.fft.fft(v_imp_filt))
v_mag = la.remove_hermitian_half(v_mag[None,:])[0]
return v_mag
def get_formant_locations_from_raw_long_frame(v_sig, v_pm, nx, fft_len):
'''
nx: frame index
'''
#v_sig, fs = la.read_audio_file(wavfile)
# Epoch detection:
#v_pm_sec, v_voi = la.reaper_epoch_detection(wavfile)
#v_pm = lu.round_to_int(v_pm_sec * fs)
# Raw-long Frame extraction:
v_frm_long = v_sig[v_pm[nx-2]:v_pm[nx+2]+1]
# Win:
left_len = v_pm[nx] - v_pm[nx-2]
right_len = v_pm[nx+2] - v_pm[nx]
v_win = la.gen_non_symmetric_win(left_len, right_len, np.hanning, b_norm=False)
v_frm_long_win = v_frm_long * v_win
# Spectrum:
v_mag = np.absolute(np.fft.fft(v_frm_long_win, n=fft_len))
v_mag_db = la.db(la.remove_hermitian_half(v_mag[None,:])[0])
# Formant extraction -LPC method:--------------------------------------------------
v_lpc, v_e, v_refl = lpc(v_frm_long_win, 120)
b_use_lpc_roots = False
if b_use_lpc_roots:
v_lpc_roots = np.roots(v_lpc)
v_lpc_angles = np.angle(v_lpc_roots)
v_lpc_angles = v_lpc_angles[v_lpc_angles>=0]
v_lpc_angles = np.sort(v_lpc_angles)
fft_len_half = 1 + fft_len / 2
v_lpc_roots_bins = v_lpc_angles * fft_len_half / np.pi
v_lpc_mag = lpc_to_mag(v_lpc, fft_len=fft_len)
v_lpc_mag_db = la.db(v_lpc_mag)
v_lpc_mag_db = v_lpc_mag_db - np.mean(v_lpc_mag_db) + np.mean(v_mag_db)
v_frmnts_bins, v_frmnts_gains_db = get_formant_locations_from_spec_env(v_lpc_mag_db)
# Getting bandwidth:
fft_len_half = 1 + fft_len / 2
v_vall_bins = get_formant_locations_from_spec_env(-v_lpc_mag_db)[0]
v_vall_bins = np.r_[0, v_vall_bins, fft_len_half-1]
nfrmnts = v_frmnts_bins.size
v_frmnts_bw = np.zeros(nfrmnts) - 1.0
for nx_f in xrange(nfrmnts):
#Left slope:
curr_frmnt_bin = v_frmnts_bins[nx_f]
curr_vall_l_bin = v_vall_bins[nx_f]
curr_vall_r_bin = v_vall_bins[nx_f+1]
curr_midp_l = int((curr_frmnt_bin + curr_vall_l_bin) / 2.0)
curr_midp_r = int((curr_frmnt_bin + curr_vall_r_bin) / 2.0)
# Protection:
if curr_midp_l==curr_frmnt_bin:
curr_midp_l = curr_vall_l_bin
if curr_midp_r==curr_frmnt_bin:
curr_midp_r = curr_vall_r_bin
#print(nx_f)
# 27 y 32
#if ((nx==73) and (nx_f==27)): import ipdb; ipdb.set_trace(context=8) # breakpoint c4f78f1e //
slope_l = (v_frmnts_gains_db[nx_f] - v_lpc_mag_db[curr_midp_l]) / (curr_frmnt_bin - curr_midp_l).astype(float)
slope_r = (v_frmnts_gains_db[nx_f] - v_lpc_mag_db[curr_midp_r]) / (curr_frmnt_bin - curr_midp_r).astype(float)
slope_ave = (slope_l - slope_r) / 2.0
v_frmnts_bw[nx_f] = 1.0 / slope_ave
# Filtering by bandwidth:
bw_thress = 7.0
v_frmnts_bins = v_frmnts_bins[v_frmnts_bw<bw_thress]
v_frmnts_gains_db = v_frmnts_gains_db[v_frmnts_bw<bw_thress]
v_frmnts_bw = v_frmnts_bw[v_frmnts_bw<bw_thress]
# Computing frame short:--------------------------------
# Win:
left_len_short = v_pm[nx] - v_pm[nx-1]
right_len_short = v_pm[nx+1] - v_pm[nx]
v_win_short = la.gen_non_symmetric_win(left_len_short, right_len_short, np.hanning, b_norm=False)
v_frm_short = v_sig[v_pm[nx-1]:v_pm[nx+1]+1]
v_frm_short_win = v_frm_short * v_win_short
shift = v_pm[nx] - v_pm[nx-1]
# Formant extraction - True envelope method:----------------------------------------
# Not finished.
#v_true_env_db = la.true_envelope(v_mag_db[None,:], in_type='db', ncoeffs=400, thres_db=0.1)[0]
if False:
plt.figure(); plt.plot(v_mag_db); plt.plot(v_lpc_mag_db); plt.grid(); plt.show()
return v_mag_db, v_lpc_mag_db, v_frmnts_bins, v_frmnts_gains_db, v_frmnts_bw, v_frm_short_win, shift
def speech_interp_with_anchors(wavfile_a, wavfile_b, nx_strt_a, nx_strt_b, nframes, fft_len):
# MagPhase analysis:
m_mag_a, m_real_a, m_imag_a, v_f0_a, fs, v_shift_a = mp.analysis_lossless(wavfile_a)
m_mag_b, m_real_b, m_imag_b, v_f0_b, fs, v_shift_b = mp.analysis_lossless(wavfile_b)
v_pm_a = la.shift_to_pm(v_shift_a)
v_pm_b = la.shift_to_pm(v_shift_b)
v_sig_a, fs = la.read_audio_file(wavfile_a)
v_sig_b, fs = la.read_audio_file(wavfile_b)
# Get formants:
v_mag_db_a_dummy, v_lpc_mag_db_a, v_frmnts_bins_a, v_frmnts_gains_db_a, v_frmnts_bw_a, v_frm_short_a, shift_a = get_formant_locations_from_raw_long_frame(v_sig_a, v_pm_a, nx_strt_a, fft_len)
v_mag_db_b_dummy, v_lpc_mag_db_b, v_frmnts_bins_b, v_frmnts_gains_db_b, v_frmnts_bw_b, v_frm_short_b, shift_b = get_formant_locations_from_raw_long_frame(v_sig_b, v_pm_b, nx_strt_a+nframes, fft_len)
# Formant mapping:----------------------------------------------------------------
v_frmnts_bins_a_filt, v_frmnts_bins_b_filt = formant_mapping(v_frmnts_bins_a, v_frmnts_gains_db_a, v_frmnts_bins_b, v_frmnts_gains_db_b, fft_len)
v_shifts_syn = np.zeros(nframes, dtype='int')
m_frms_syn = np.zeros((nframes, fft_len))
for nx_frm in xrange(nframes):
sp_weight = nx_frm / (nframes-1.0)
nx_a = nx_strt_a + nx_frm
nx_b = nx_strt_b + nx_frm
# Computing mag spectrum:
v_mag_db_a = get_formant_locations_from_raw_long_frame(v_sig_a, v_pm_a, nx_a, fft_len)[0]
v_mag_db_b = get_formant_locations_from_raw_long_frame(v_sig_b, v_pm_b, nx_b, fft_len)[0]
# NOT FINISHED !!
# Warping:---------------------------------------------------------------------
# True envelope:
v_true_env_db_a = la.true_envelope(v_mag_db_a[None,:], in_type='db', ncoeffs=400, thres_db=0.1)[0]
v_true_env_db_b = la.true_envelope(v_mag_db_b[None,:], in_type='db', ncoeffs=400, thres_db=0.1)[0]
v_sp_env_db_a_warp = warp_mag_spec(v_true_env_db_a, v_frmnts_bins_a_filt, v_frmnts_bins_b_filt, fft_len, sp_weight)
v_sp_env_db_b_warp = warp_mag_spec(v_true_env_db_b, v_frmnts_bins_b_filt, v_frmnts_bins_a_filt, fft_len, (1-sp_weight))
#v_sp_env_db_a_warp = warp_mag_spec(v_lpc_mag_db_a, v_frmnts_bins_a_filt, v_frmnts_bins_b_filt, fft_len, sp_weight)
#v_sp_env_db_b_warp = warp_mag_spec(v_lpc_mag_db_b, v_frmnts_bins_b_filt, v_frmnts_bins_a_filt, fft_len, (1-sp_weight))
# Spectral envelope mix:-------------------------------------------------------
v_sp_env_db_targ = v_sp_env_db_a_warp * (1.0-sp_weight) + v_sp_env_db_b_warp * sp_weight
# Source mix:------------------------------------------------------------------
v_spec_diff_db_a = v_sp_env_db_targ - v_true_env_db_a
v_spec_diff_db_b = v_sp_env_db_targ - v_true_env_db_b
# Filtering (FFT filter):
v_frm_short_a_ext_filt = fft_filter(v_frm_short_a, shift_a, v_spec_diff_db_a, fft_len)
v_frm_short_b_ext_filt = fft_filter(v_frm_short_b, shift_b, v_spec_diff_db_b, fft_len)
# Mix signal:
v_frm_short_ext_filt = v_frm_short_a_ext_filt * (1.0-sp_weight) + v_frm_short_b_ext_filt * sp_weight
#v_frm_short_ext_filt = v_frm_short_a_ext_filt
# Mix shifts:
shift_mix = lu.round_to_int(shift_a * (1.0-sp_weight) + shift_b * sp_weight)
# Save:
v_shifts_syn[nx_frm] = shift_mix
m_frms_syn[nx_frm, :] = v_frm_short_ext_filt
if False:
plt.figure(); plt.plot(v_frm_short_a_ext_filt); plt.plot(v_frm_short_b_ext_filt); plt.grid(); plt.show()
plt.figure(); plt.plot(v_frm_short_a_ext_filt); plt.plot(v_frm_short_b_ext_filt); plt.plot(v_frm_short_ext_filt); plt.grid(); plt.show()
# Merge:
m_frms_syn_dc = np.fft.fftshift(m_frms_syn, axes=1)
m_fft_syn = la.remove_hermitian_half(np.fft.fft(m_frms_syn_dc))
m_mag_syn, m_real_syn, m_imag_syn = compute_lossless_spec_feats(m_fft_syn)
m_mag_merged = np.vstack((m_mag_a[:nx_strt_a,:] , m_mag_syn , m_mag_b[(nx_strt_b+nframes):,:]))
m_real_merged = np.vstack((m_real_a[:nx_strt_a,:] , m_real_syn , m_real_b[(nx_strt_b+nframes):,:]))
m_imag_merged = np.vstack((m_imag_a[:nx_strt_a,:] , m_imag_syn , m_imag_b[(nx_strt_b+nframes):,:]))
v_shift_merged = np.r_[ v_shift_a[:nx_strt_a] , v_shifts_syn , v_shift_b[(nx_strt_b+nframes):] ]
v_sig_merged = synthesis_from_lossless(m_mag_merged, m_real_merged, m_imag_merged, v_shift_merged)
return v_sig_merged, fs
def get_formant_locations_from_raw_long_frame(v_sig, v_pm, nx, fft_len):
'''
nx: frame index
'''
#v_sig, fs = la.read_audio_file(wavfile)
# Epoch detection:
#v_pm_sec, v_voi = la.reaper_epoch_detection(wavfile)
#v_pm = lu.round_to_int(v_pm_sec * fs)
# Raw-long Frame extraction:
v_frm_long = v_sig[v_pm[nx-2]:v_pm[nx+2]+1]
# Win:
left_len = v_pm[nx] - v_pm[nx-2]
right_len = v_pm[nx+2] - v_pm[nx]
v_win = la.gen_non_symmetric_win(left_len, right_len, np.hanning, b_norm=False)
v_frm_long_win = v_frm_long * v_win
# Spectrum:
v_mag = la.remove_hermitian_half(np.absolute(np.fft.fft(v_frm_long_win, n=fft_len)[None,:]))[0]
v_mag_db = la.db(v_mag)
# Mel warping:
alpha = 0.50 # 0.55 - 0.60
#ncoeffs = 2048 # must be even
v_mag_mel = la.sp_mel_warp(v_mag[None,:], fft_len/2, alpha=alpha, in_type=3)[0]
v_sp_cmplx = la.build_min_phase_from_mag_spec(v_mag_mel[None,:])[0]
v_sp_cmplx_ext = la.add_hermitian_half(v_sp_cmplx[None,:], data_type='complex')[0]
v_frm_long_win_mel = np.fft.ifft(v_sp_cmplx_ext).real
if False:
plt.close('all')
pl(la.db(v_mag))
pl(la.db(np.absolute(v_sp_cmplx)))
pl(v_frm_long_win_mel)
# Formant extraction -LPC method:--------------------------------------------------
n_lpc_coeffs = 30 # 40
v_lpc_mel, v_e, v_refl = lpc(v_frm_long_win_mel, n_lpc_coeffs)
v_lpc_mag_mel = lpc_to_mag(v_lpc_mel, fft_len=fft_len)
v_lpc_mag_mel_db = la.db(v_lpc_mag_mel)
v_lpc_mag_mel_db = v_lpc_mag_mel_db - np.mean(v_lpc_mag_mel_db) + np.mean(la.db(v_mag_mel))
v_frmnts_bins_mel, v_frmnts_gains_db = get_formant_locations_from_spec_env(v_lpc_mag_mel_db)
# Getting bandwidth:
fft_len_half = 1 + fft_len / 2
v_vall_bins = get_formant_locations_from_spec_env(-v_lpc_mag_mel_db)[0]
v_vall_bins = np.r_[0, v_vall_bins, fft_len_half-1]
nfrmnts = v_frmnts_bins_mel.size
v_frmnts_bw_mel = np.zeros(nfrmnts) - 1.0
for nx_f in xrange(nfrmnts):
#Left slope:
curr_frmnt_bin = v_frmnts_bins_mel[nx_f]
curr_vall_l_bin = v_vall_bins[nx_f]
curr_vall_r_bin = v_vall_bins[nx_f+1]
curr_midp_l = int((curr_frmnt_bin + curr_vall_l_bin) / 2.0)
curr_midp_r = int((curr_frmnt_bin + curr_vall_r_bin) / 2.0)
# Protection:
if curr_midp_l==curr_frmnt_bin:
curr_midp_l = curr_vall_l_bin
if curr_midp_r==curr_frmnt_bin:
curr_midp_r = curr_vall_r_bin
#print(nx_f)
# 27 y 32
slope_l = (v_frmnts_gains_db[nx_f] - v_lpc_mag_mel_db[curr_midp_l]) / (curr_frmnt_bin - curr_midp_l).astype(float)
slope_r = (v_frmnts_gains_db[nx_f] - v_lpc_mag_mel_db[curr_midp_r]) / (curr_frmnt_bin - curr_midp_r).astype(float)
slope_ave = (slope_l - slope_r) / 2.0
v_frmnts_bw_mel[nx_f] = 1.0 / slope_ave
# Filtering by bandwidth:
# bw_thress = 7.0
# v_frmnts_bins_mel = v_frmnts_bins_mel[v_frmnts_bw_mel<bw_thress]
# v_frmnts_gains_db = v_frmnts_gains_db[v_frmnts_bw_mel<bw_thress]
# v_frmnts_bw_mel = v_frmnts_bw_mel[v_frmnts_bw_mel<bw_thress]
# Computing frame short:--------------------------------
# Win:
left_len_short = v_pm[nx] - v_pm[nx-1]
right_len_short = v_pm[nx+1] - v_pm[nx]
v_win_short = la.gen_non_symmetric_win(left_len_short, right_len_short, np.hanning, b_norm=False)
v_frm_short = v_sig[v_pm[nx-1]:v_pm[nx+1]+1]
v_frm_short_win = v_frm_short * v_win_short
shift = v_pm[nx] - v_pm[nx-1]
# Formant extraction - True envelope method:----------------------------------------
# Not finished.
#v_true_env_db = la.true_envelope(v_mag_db[None,:], in_type='db', ncoeffs=400, thres_db=0.1)[0]
if True:
plt.figure(); plt.plot(la.db(v_mag_mel)); plt.plot(v_lpc_mag_mel_db); plt.grid(); plt.show()
#pl(v_mag_db)
if True: import ipdb; ipdb.set_trace(context=8) # breakpoint 906d26d6 //
return v_mag_db, v_lpc_mag_mel_db, v_frmnts_bins_mel, v_frmnts_gains_db, v_frmnts_bw_mel, v_frm_short_win, shift