2023-01-18 10:49:58 +01:00

661 lines
18 KiB
Python

#
# Copyright 2022 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import numpy as np
import scipy.signal as signal
import lc3
import tables as T, appendix_c as C
### ------------------------------------------------------------------------ ###
class Resampler_12k8:
def __init__(self, dt, sr, history = 0):
self.sr = sr
self.p = 192 // T.SRATE_KHZ[sr]
self.w = 240 // self.p
self.n = ((T.DT_MS[dt] * 128) / 10).astype(int)
self.d = [ 44, 24 ][dt]
self.x = np.zeros(self.w + T.NS[dt][sr])
self.u = np.zeros(self.n + 2)
self.y = np.zeros(self.n + self.d + history)
def resample(self, x):
p = self.p
w = self.w
d = self.d
n = self.n
### Sliding window
self.x[:w] = self.x[-w:]
self.x[w:] = x
self.u[:2] = self.u[-2:]
if len(self.y) > 2*n + d:
self.y[n+d:-n] = self.y[d+2*n:]
if len(self.y) > n + d:
self.y[-n:] = self.y[:n]
self.y[:d] = self.y[n:d+n]
x = self.x
u = self.u
### 3.3.9.3 Resampling
h = np.zeros(240 + p)
h[-119:] = T.LTPF_H12K8[:119]
h[ :120] = T.LTPF_H12K8[119:]
for i in range(n):
e = (15 * i) // p
f = (15 * i) % p
k = np.arange(-120, 120 + p, p) - f
u[2+i] = p * np.dot( x[e:e+w+1], np.take(h, k) )
if self.sr == T.SRATE_8K:
u = 0.5 * u
### 3.3.9.4 High-pass filtering
b = [ 0.9827947082978771, -1.9655894165957540, 0.9827947082978771 ]
a = [ 1 , -1.9652933726226904, 0.9658854605688177 ]
self.y[d:d+n] = b[0] * u[2:] + b[1] * u[1:-1] + b[2] * u[:-2]
for i in range(n):
self.y[d+i] -= a[1] * self.y[d+i-1] + a[2] * self.y[d+i-2]
return self.y
class Resampler_6k4:
def __init__(self, n, history = 0):
self.x = np.zeros(n + 5)
self.n = n // 2
self.y = np.zeros(self.n + history)
def resample(self, x):
n = self.n
### Sliding window
self.x[:3] = self.x[-5:-2]
self.x[3:] = x[:2*n+2]
x = self.x
if len(self.y) > 2*n:
self.y[n:-n] = self.y[2*n:]
if len(self.y) > n:
self.y[-n:] = self.y[:n]
### 3.3.9.5 Downsampling to 6.4 KHz
h = [ 0.1236796411180537, 0.2353512128364889, 0.2819382920909148,
0.2353512128364889, 0.1236796411180537 ]
self.y[:n] = [ np.dot(x[2*i:2*i+5], h) for i in range(self.n) ]
return self.y
def initial_hp50_state():
return { 's1': 0, 's2': 0 }
### ------------------------------------------------------------------------ ###
class Ltpf:
def __init__(self, dt, sr):
self.dt = dt
self.sr = sr
(self.pitch_present, self.pitch_index) = (None, None)
class LtpfAnalysis(Ltpf):
def __init__(self, dt, sr):
super().__init__(dt, sr)
self.resampler_12k8 = Resampler_12k8(
dt, sr, history = 232)
self.resampler_6k4 = Resampler_6k4(
self.resampler_12k8.n, history = 114)
self.active = False
self.tc = 0
self.pitch = 0
self.nc = np.zeros(2)
def get_data(self):
return { 'active' : self.active,
'pitch_index' : self.pitch_index }
def get_nbits(self):
return 1 + 10 * int(self.pitch_present)
def correlate(self, x, n, k0, k1):
return [ np.dot(x[:n], np.take(x, np.arange(n) - k)) \
for k in range(k0, 1+k1) ]
def norm_corr(self, x, n, k):
u = x[:n]
v = np.take(x, np.arange(n) - k)
uv = np.dot(u, v)
return uv / np.sqrt(np.dot(u, u) * np.dot(v, v)) if uv > 0 else 0
def run(self, x):
### 3.3.9.3-4 Resampling
x_12k8 = self.resampler_12k8.resample(x)
### 3.3.9.5-6 Pitch detection algorithm
x = self.resampler_6k4.resample(x_12k8)
n = self.resampler_6k4.n
r = self.correlate(x, n, 17, 114)
rw = r * (1 - 0.5 * np.arange(len(r)) / (len(r) - 1))
tc = self.tc
k0 = max(0, tc-4)
k1 = min(len(r)-1, tc+4)
t = [ 17 + np.argmax(rw), 17 + k0 + np.argmax(r[k0:1+k1]) ]
nc = [ self.norm_corr(x, n, t[i]) for i in range(2) ]
ti = int(nc[1] > 0.85 * nc[0])
self.tc = t[ti] - 17
self.pitch_present = bool(nc[ti] > 0.6)
### 3.3.9.7 Pitch-lag parameter
if self.pitch_present:
tc = self.tc + 17
x = x_12k8
n = self.resampler_12k8.n
k0 = max( 32, 2*tc-4)
k1 = min(228, 2*tc+4)
r = self.correlate(x, n, k0-4, k1+4)
e = k0 + np.argmax(r[4:-4])
h = np.zeros(42)
h[-15:] = T.LTPF_H4[:15]
h[ :16] = T.LTPF_H4[15:]
m = np.arange(-4, 5)
s = [ np.dot( np.take(r, e-k0+4 + m), np.take(h, 4*m-d) ) \
for d in range(-3, 4) ]
f = np.argmax(s[3:]) if e <= 32 else \
-3 + np.argmax(s) if e < 127 else \
-2 + 2*np.argmax(s[1:-1:2]) if e < 157 else 0
e -= (f < 0)
f += 4*(f < 0)
self.pitch_index = 4*e + f - 128 if e < 127 else \
2*e + f//2 + 126 if e < 157 else e + 283
else:
e = f = 0
self.pitch_index = 0
### 3.3.9.8 Activation bit
h = np.zeros(24)
h[-7:] = T.LTPF_HI[:7]
h[ :8] = T.LTPF_HI[7:]
k = np.arange(-2, 3)
u = [ np.dot( np.take(x, i-k), np.take(h, 4*k) ) \
for i in range(n) ]
v = [ np.dot( np.take(x, i-k), np.take(h, 4*k-f) ) \
for i in range(-e, n-e) ]
nc = max(0, np.dot(u, v)) / np.sqrt(np.dot(u, u) * np.dot(v, v)) \
if self.pitch_present else 0
pitch = e + f/4
if not self.active:
active = (self.dt == T.DT_10M or self.nc[1] > 0.94) \
and self.nc[0] > 0.94 and nc > 0.94
else:
dp = abs(pitch - self.pitch)
dc = nc - self.nc[0]
active = nc > 0.9 or (dp < 2 and dc > -0.1 and nc > 0.84)
if not self.pitch_present:
active = False
pitch = 0
nc = 0
self.active = active
self.pitch = pitch
self.nc[1] = self.nc[0]
self.nc[0] = nc
return self.pitch_present
def disable(self):
self.active = False
def store(self, b):
b.write_uint(self.active, 1)
b.write_uint(self.pitch_index, 9)
class LtpfSynthesis(Ltpf):
C_N = [ T.LTPF_N_8K , T.LTPF_N_16K,
T.LTPF_N_24K, T.LTPF_N_32K, T.LTPF_N_48K ]
C_D = [ T.LTPF_D_8K , T.LTPF_D_16K,
T.LTPF_D_24K, T.LTPF_D_32K, T.LTPF_D_48K ]
def __init__(self, dt, sr):
super().__init__(dt, sr)
self.C_N = LtpfSynthesis.C_N[sr]
self.C_D = LtpfSynthesis.C_D[sr]
ns = T.NS[dt][sr]
self.active = [ False, False ]
self.pitch_index = 0
max_pitch_12k8 = 228
max_pitch = max_pitch_12k8 * T.SRATE_KHZ[self.sr] / 12.8
max_pitch = np.ceil(max_pitch).astype(int)
self.x = np.zeros(ns)
self.y = np.zeros(max_pitch + len(self.C_D[0]))
self.p_e = [ 0, 0 ]
self.p_f = [ 0, 0 ]
self.c_n = [ None, None ]
self.c_d = [ None, None ]
def load(self, b):
self.active[0] = bool(b.read_uint(1))
self.pitch_index = b.read_uint(9)
def disable(self):
self.active[0] = False
self.pitch_index = 0
def run(self, x, nbytes):
sr = self.sr
dt = self.dt
### 3.4.9.4 Filter parameters
pitch_index = self.pitch_index
if pitch_index >= 440:
p_e = pitch_index - 283
p_f = 0
elif pitch_index >= 380:
p_e = pitch_index // 2 - 63
p_f = 2*(pitch_index - 2*(p_e + 63))
else:
p_e = pitch_index // 4 + 32
p_f = pitch_index - 4*(p_e - 32)
p = (p_e + p_f / 4) * T.SRATE_KHZ[self.sr] / 12.8
self.p_e[0] = int(p * 4 + 0.5) // 4
self.p_f[0] = int(p * 4 + 0.5) - 4*self.p_e[0]
nbits = round(nbytes*80 / T.DT_MS[dt])
g_idx = max(nbits // 80, 3+sr) - (3+sr)
g = [ 0.4, 0.35, 0.3, 0.25 ][g_idx] if g_idx < 4 else 0
g_idx = min(g_idx, 3)
self.c_n[0] = 0.85 * g * LtpfSynthesis.C_N[sr][g_idx]
self.c_d[0] = g * LtpfSynthesis.C_D[sr][self.p_f[0]]
### 3.4.9.2 Transition handling
n0 = (T.SRATE_KHZ[sr] * 1000) // 400
ns = T.NS[dt][sr]
x = np.append(x, self.x)
y = np.append(np.zeros(ns), self.y)
yc = y.copy()
c_n = self.c_n
c_d = self.c_d
l_n = len(c_n[0])
l_d = len(c_d[0])
d = [ self.p_e[0] - (l_d - 1) // 2,
self.p_e[1] - (l_d - 1) // 2 ]
for k in range(n0):
if not self.active[0] and not self.active[1]:
y[k] = x[k]
elif self.active[0] and not self.active[1]:
u = np.dot(c_n[0], np.take(x, k - np.arange(l_n))) - \
np.dot(c_d[0], np.take(y, k - d[0] - np.arange(l_d)))
y[k] = x[k] - (k/n0) * u
elif not self.active[0] and self.active[1]:
u = np.dot(c_n[1], np.take(x, k - np.arange(l_n))) - \
np.dot(c_d[1], np.take(y, k - d[1] - np.arange(l_d)))
y[k] = x[k] - (1 - k/n0) * u
elif self.p_e[0] == self.p_e[1] and self.p_f[0] == self.p_f[1]:
u = np.dot(c_n[0], np.take(x, k - np.arange(l_n))) - \
np.dot(c_d[0], np.take(y, k - d[0] - np.arange(l_d)))
y[k] = x[k] - u
else:
u = np.dot(c_n[1], np.take(x, k - np.arange(l_n))) - \
np.dot(c_d[1], np.take(y, k - d[1] - np.arange(l_d)))
yc[k] = x[k] - (1 - k/n0) * u
u = np.dot(c_n[0], np.take(yc, k - np.arange(l_n))) - \
np.dot(c_d[0], np.take(y , k - d[0] - np.arange(l_d)))
y[k] = yc[k] - (k/n0) * u
### 3.4.9.3 Remainder of the frame
for k in range(n0, ns):
if not self.active[0]:
y[k] = x[k]
else:
u = np.dot(c_n[0], np.take(x, k - np.arange(l_n))) - \
np.dot(c_d[0], np.take(y, k - d[0] - np.arange(l_d)))
y[k] = x[k] - u
### Sliding window
self.active[1] = self.active[0]
self.p_e[1] = self.p_e[0]
self.p_f[1] = self.p_f[0]
self.c_n[1] = self.c_n[0]
self.c_d[1] = self.c_d[0]
self.x = x[:ns]
self.y = np.append(self.y[ns:], y[:ns])
return y[:ns]
def initial_state():
return { 'active' : False, 'pitch': 0, 'nc': np.zeros(2),
'hp50' : initial_hp50_state(),
'x_12k8' : np.zeros(384), 'x_6k4' : np.zeros(178), 'tc' : 0 }
def initial_sstate():
return { 'active': False, 'pitch': 0,
'c': np.zeros(2*12), 'x': np.zeros(12) }
### ------------------------------------------------------------------------ ###
def check_resampler(rng, dt, sr):
ns = T.NS[dt][sr]
nt = (5 * T.SRATE_KHZ[sr]) // 4
ok = True
r = Resampler_12k8(dt, sr)
hp50_c = initial_hp50_state()
x_c = np.zeros(nt)
y_c = np.zeros(384)
for run in range(10):
x = ((2 * rng.random(ns)) - 1) * (2 ** 15 - 1)
y = r.resample(x)
x_c = np.append(x_c[-nt:], x.astype(np.int16))
y_c[:-r.n] = y_c[r.n:]
y_c = lc3.ltpf_resample(dt, sr, hp50_c, x_c, y_c)
ok = ok and np.amax(np.abs(y_c[-r.d-r.n:] - y[:r.d+r.n]/2)) < 4
return ok
def check_resampler_appendix_c(dt):
sr = T.SRATE_16K
ok = True
nt = (5 * T.SRATE_KHZ[sr]) // 4
n = [ 96, 128 ][dt]
k = [ 44, 24 ][dt] + n
state = initial_hp50_state()
x = np.append(np.zeros(nt), C.X_PCM[dt][0])
y = np.zeros(384)
y = lc3.ltpf_resample(dt, sr, state, x, y)
u = y[-k:len(C.X_TILDE_12K8D[dt][0])-k]
ok = ok and np.amax(np.abs(u - C.X_TILDE_12K8D[dt][0]/2)) < 2
x = np.append(x[-nt:], C.X_PCM[dt][1])
y[:-n] = y[n:]
y = lc3.ltpf_resample(dt, sr, state, x, y)
u = y[-k:len(C.X_TILDE_12K8D[dt][1])-k]
ok = ok and np.amax(np.abs(u - C.X_TILDE_12K8D[dt][1]/2)) < 2
return ok
def check_analysis(rng, dt, sr):
ns = T.NS[dt][sr]
nt = (5 * T.SRATE_KHZ[sr]) // 4
ok = True
state_c = initial_state()
x_c = np.zeros(ns+nt)
ltpf = LtpfAnalysis(dt, sr)
t = np.arange(100 * ns) / (T.SRATE_KHZ[sr] * 1000)
s = signal.chirp(t, f0=10, f1=3e3, t1=t[-1], method='logarithmic')
for i in range(20):
x = s[i*ns:(i+1)*ns] * (2 ** 15 - 1)
pitch_present = ltpf.run(x)
data = ltpf.get_data()
x_c = np.append(x_c[-nt:], x.astype(np.int16))
(pitch_present_c, data_c) = lc3.ltpf_analyse(dt, sr, state_c, x_c)
ok = ok and (not pitch_present or state_c['tc'] == ltpf.tc)
ok = ok and np.amax(np.abs(state_c['nc'][0] - ltpf.nc[0])) < 1e-2
ok = ok and pitch_present_c == pitch_present
ok = ok and data_c['active'] == data['active']
ok = ok and data_c['pitch_index'] == data['pitch_index']
ok = ok and lc3.ltpf_get_nbits(pitch_present) == ltpf.get_nbits()
return ok
def check_synthesis(rng, dt, sr):
ok = True
ns = T.NS[dt][sr]
nd = 18 * T.SRATE_KHZ[sr]
synthesis = LtpfSynthesis(dt, sr)
state_c = initial_sstate()
x_c = np.zeros(nd+ns)
for i in range(50):
pitch_present = bool(rng.integers(0, 10) >= 1)
if not pitch_present:
synthesis.disable()
else:
synthesis.active[0] = bool(rng.integers(0, 5) >= 1)
synthesis.pitch_index = rng.integers(0, 512)
data_c = None if not pitch_present else \
{ 'active' : synthesis.active[0],
'pitch_index' : synthesis.pitch_index }
x = rng.random(ns) * 1e4
nbytes = rng.integers(10*(2+sr), 10*(6+sr))
x_c[:nd] = x_c[ns:]
x_c[nd:] = x
y = synthesis.run(x, nbytes)
x_c = lc3.ltpf_synthesize(dt, sr, nbytes, state_c, data_c, x_c)
ok = ok and np.amax(np.abs(x_c[nd:] - y)) < 1e-2
return ok
def check_analysis_appendix_c(dt):
sr = T.SRATE_16K
nt = (5 * T.SRATE_KHZ[sr]) // 4
ok = True
state = initial_state()
x = np.append(np.zeros(nt), C.X_PCM[dt][0])
(pitch_present, data) = lc3.ltpf_analyse(dt, sr, state, x)
ok = ok and C.T_CURR[dt][0] - state['tc'] == 17
ok = ok and np.amax(np.abs(state['nc'][0] - C.NC_LTPF[dt][0])) < 1e-5
ok = ok and pitch_present == C.PITCH_PRESENT[dt][0]
ok = ok and data['pitch_index'] == C.PITCH_INDEX[dt][0]
ok = ok and data['active'] == C.LTPF_ACTIVE[dt][0]
x = np.append(x[-nt:], C.X_PCM[dt][1])
(pitch_present, data) = lc3.ltpf_analyse(dt, sr, state, x)
ok = ok and C.T_CURR[dt][1] - state['tc'] == 17
ok = ok and np.amax(np.abs(state['nc'][0] - C.NC_LTPF[dt][1])) < 1e-5
ok = ok and pitch_present == C.PITCH_PRESENT[dt][1]
ok = ok and data['pitch_index'] == C.PITCH_INDEX[dt][1]
ok = ok and data['active'] == C.LTPF_ACTIVE[dt][1]
return ok
def check_synthesis_appendix_c(dt):
sr = T.SRATE_16K
ok = True
if dt != T.DT_10M:
return ok
ns = T.NS[dt][sr]
nd = 18 * T.SRATE_KHZ[sr]
NBYTES = [ C.LTPF_C2_NBITS // 8, C.LTPF_C3_NBITS // 8,
C.LTPF_C4_NBITS // 8, C.LTPF_C5_NBITS // 8 ]
ACTIVE = [ C.LTPF_C2_ACTIVE, C.LTPF_C3_ACTIVE,
C.LTPF_C4_ACTIVE, C.LTPF_C5_ACTIVE ]
PITCH_INDEX = [ C.LTPF_C2_PITCH_INDEX, C.LTPF_C3_PITCH_INDEX,
C.LTPF_C4_PITCH_INDEX, C.LTPF_C5_PITCH_INDEX ]
X = [ C.LTPF_C2_X, C.LTPF_C3_X,
C.LTPF_C4_X, C.LTPF_C5_X ]
PREV = [ C.LTPF_C2_PREV, C.LTPF_C3_PREV,
C.LTPF_C4_PREV, C.LTPF_C5_PREV ]
TRANS = [ C.LTPF_C2_TRANS, C.LTPF_C3_TRANS,
C.LTPF_C4_TRANS, C.LTPF_C5_TRANS ]
for i in range(4):
state = initial_sstate()
nbytes = NBYTES[i]
data = { 'active' : ACTIVE[i][0], 'pitch_index' : PITCH_INDEX[i][0] }
x = np.append(np.zeros(nd), X[i][0])
lc3.ltpf_synthesize(dt, sr, nbytes, state, data, x)
data = { 'active' : ACTIVE[i][1], 'pitch_index' : PITCH_INDEX[i][1] }
x[ :nd-ns] = PREV[i][0][-nd+ns:]
x[nd-ns:nd] = PREV[i][1]
x[nd:nd+ns] = X[i][1]
y = lc3.ltpf_synthesize(dt, sr, nbytes, state, data, x)[nd:]
ok = ok and np.amax(np.abs(y - TRANS[i])) < 1e-3
return ok
def check():
rng = np.random.default_rng(1234)
ok = True
for dt in range(T.NUM_DT):
for sr in range(T.NUM_SRATE):
ok = ok and check_resampler(rng, dt, sr)
ok = ok and check_analysis(rng, dt, sr)
ok = ok and check_synthesis(rng, dt, sr)
for dt in range(T.NUM_DT):
ok = ok and check_resampler_appendix_c(dt)
ok = ok and check_analysis_appendix_c(dt)
ok = ok and check_synthesis_appendix_c(dt)
return ok
### ------------------------------------------------------------------------ ###