/*
* E_MAIN_init
*
* Parameters:
* spe_state I/O: pointer to state structure
*
* Function:
* Initialisation of variables for the coder section.
* Memory allocation.
*
* Returns:
* void
*/
Word16 E_MAIN_init(void **spe_state)
{
Coder_State *st;
*spe_state = NULL;
/* allocate memory */
if ((st = (Coder_State *) malloc(sizeof(Coder_State))) == NULL)
{
return(-1);
}
st->vadSt = NULL;
st->dtx_encSt = NULL;
E_DTX_init(&(st->dtx_encSt));
E_DTX_vad_init(&(st->vadSt));
E_MAIN_reset((void *) st, 1);
*spe_state = (void*)st;
return(0);
}
/*
* E_MAIN_encode
*
* Parameters:
* mode I: used mode
* input_sp I: 320 new speech samples (at 16 kHz)
* prms O: output parameters
* spe_state B: state structure
* allow_dtx I: DTX ON/OFF
*
* Function:
* Main coder routine.
*
* Returns:
* void
*/
Word16 E_MAIN_encode(Word16 * mode, Word16 speech16k[], Word16 prms[],
void *spe_state, Word16 allow_dtx)
{
/* Float32 */
Float32 f_speech16k[L_FRAME16k]; /* Speech vector */
Float32 f_old_exc[(L_FRAME + 1) + PIT_MAX + L_INTERPOL]; /* Excitation vector */
Float32 f_exc2[L_FRAME]; /* excitation vector */
Float32 error[M + L_SUBFR]; /* error of quantization */
Float32 A[NB_SUBFR * (M + 1)]; /* A(z) unquantized for the 4 subframes */
Float32 Aq[NB_SUBFR * (M + 1)]; /* A(z) quantized for the 4 subframes */
Float32 xn[L_SUBFR]; /* Target vector for pitch search */
Float32 xn2[L_SUBFR]; /* Target vector for codebook search */
Float32 dn[L_SUBFR]; /* Correlation between xn2 and h1 */
Float32 cn[L_SUBFR]; /* Target vector in residual domain */
Float32 h1[L_SUBFR]; /* Impulse response vector */
Float32 f_code[L_SUBFR]; /* Fixed codebook excitation */
Float32 y1[L_SUBFR]; /* Filtered adaptive excitation */
Float32 y2[L_SUBFR]; /* Filtered adaptive excitation */
Float32 synth[L_SUBFR]; /* 12.8kHz synthesis vector */
Float32 r[M + 1]; /* Autocorrelations of windowed speech */
Float32 Ap[M + 1]; /* A(z) with spectral expansion */
Float32 ispnew[M]; /* immittance spectral pairs at 4nd sfr */
Float32 isf[M]; /* ISF (frequency domain) at 4nd sfr */
Float32 g_coeff[5], g_coeff2[2]; /* Correlations */
Float32 gain_pit;
Float32 f_tmp, gain1, gain2;
Float32 stab_fac = 0.0F, fac;
Float32 *new_speech, *speech; /* Speech vector */
Float32 *wsp; /* Weighted speech vector */
Float32 *f_exc; /* Excitation vector */
Float32 *p_A, *p_Aq; /* ptr to A(z) for the 4 subframes */
Float32 *f_pt_tmp;
/* Word32 */
Word32 indice[8]; /* quantization indices */
Word32 vad_flag, clip_gain;
Word32 T_op, T_op2, T0, T0_frac;
Word32 T0_min, T0_max;
Word32 voice_fac, Q_new = 0;
Word32 L_gain_code, l_tmp;
Word32 i, i_subfr, pit_flag;
/* Word16 */
Word16 exc2[L_FRAME]; /* excitation vector */
Word16 s_Aq[NB_SUBFR * (M + 1)]; /* A(z) quantized for the 4 subframes */
Word16 s_code[L_SUBFR]; /* Fixed codebook excitation */
Word16 ispnew_q[M]; /* quantized ISPs at 4nd subframe */
Word16 isfq[M]; /* quantized ISPs */
Word16 select, codec_mode;
Word16 index;
Word16 s_gain_pit, gain_code;
Word16 s_tmp, s_max;
Word16 corr_gain;
Word16 *exc; /* Excitation vector */
/* Other */
Coder_State *st; /* Coder states */
/* Memory Usage eval */
st = (Coder_State *)spe_state;
codec_mode = *mode;
/*
* Initialize pointers to speech vector.
*
*
* |-------|-------|-------|-------|-------|-------|
* past sp sf1 sf2 sf3 sf4 L_NEXT
* <------- Total speech buffer (L_TOTAL) ------>
* old_speech
* <------- LPC analysis window (L_WINDOW) ------>
* <-- present frame (L_FRAME) ---->
* | <----- new speech (L_FRAME) ---->
* | |
* speech |
* new_speech
*/
new_speech = st->mem_speech + L_TOTAL - L_FRAME - L_FILT + 460; /* New speech */
speech = st->mem_speech + L_TOTAL - L_FRAME - L_NEXT; /* Present frame */
exc = st->mem_exc + PIT_MAX + L_INTERPOL;
f_exc = f_old_exc + PIT_MAX + L_INTERPOL;
wsp = st->mem_wsp + (PIT_MAX / OPL_DECIM);
for(i = 0; i < L_FRAME16k; i++)
{
f_speech16k[i] = (Float32)speech16k[i];
}
Q_new = -st->mem_q;
for(i = 0; i < (PIT_MAX + L_INTERPOL); i++)
{
f_old_exc[i] = (Float32)(st->mem_exc[i] * pow(2, Q_new));
}
/*
* Down sampling signal from 16kHz to 12.8kHz
*/
E_UTIL_decim_12k8(f_speech16k, L_FRAME16k, new_speech, st->mem_decim);
/* decimate with zero-padding to avoid delay of filter */
memcpy(f_code, st->mem_decim, 2 * L_FILT16k * sizeof(Float32));
memset(error, 0, L_FILT16k * sizeof(Float32));
E_UTIL_decim_12k8(error, L_FILT16k, new_speech + L_FRAME, f_code);
/*
* Perform 50Hz HP filtering of input signal.
* Perform fixed preemphasis through 1 - g z^-1
*/
E_UTIL_hp50_12k8(new_speech, L_FRAME, st->mem_sig_in);
memcpy(f_code, st->mem_sig_in, 4 * sizeof(Float32) );
E_UTIL_hp50_12k8(new_speech + L_FRAME, L_FILT, f_code);
E_UTIL_f_preemph(new_speech, PREEMPH_FAC, L_FRAME, &(st->mem_preemph));
/* last L_FILT samples for autocorrelation window */
f_tmp = st->mem_preemph;
E_UTIL_f_preemph(new_speech + L_FRAME, PREEMPH_FAC, L_FILT, &f_tmp);
/*
* Call VAD
* Preemphesis scale down signal in low frequency and keep dynamic in HF.
* Vad work slightly in future (new_speech = speech + L_NEXT - L_FILT).
*/
vad_flag = E_DTX_vad(st->vadSt, speech + L_NEXT - L_FILT);
if (vad_flag == 0)
{
st->mem_vad_hist = 1;
}
else
{
st->mem_vad_hist = 0;
}
/* DTX processing */
if (allow_dtx)
{
/* Note that mode may change here */
E_DTX_tx_handler(st->dtx_encSt, vad_flag, mode);
}
else
{
E_DTX_reset(st->dtx_encSt);
}
if(*mode != MRDTX)
{
E_MAIN_parm_store(vad_flag, &prms);
}
/*
* Perform LPC analysis
* --------------------
* - autocorrelation + lag windowing
* - Levinson-durbin algorithm to find a[]
* - convert a[] to isp[]
* - convert isp[] to isf[] for quantization
* - quantize and code the isf[]
* - convert isf[] to isp[] for interpolation
* - find the interpolated isps and convert to a[] for the 4 subframes
*/
/* LP analysis centered at 3nd surame */
E_UTIL_autocorr(st->mem_speech, r);
E_LPC_lag_wind(r + 1, M); /* Lagindowing */
E_LPC_lev_dur(A, r, M);
E_LPC_a_isp_conversion(A, ispnew, st->mem_isp, M); /* From A(z) to isp */
/* Find the interpolated isps and convert to a[] for all subframes */
E_LPC_f_int_isp_find(st->mem_isp, ispnew, A, NB_SUBFR, M);
/* update isp memory for the next frame */
memcpy(st->mem_isp, ispnew, M * sizeof(Float32));
/* Convert isps to frequency domain 0..6400 */
E_LPC_isp_isf_conversion(ispnew, isf, M);
/* check resonance for pitch clipping algorithm */
E_GAIN_clip_isf_test(isf, st->mem_gp_clip);
/*
* Perform PITCH_OL analysis
* -------------------------
* - Find the residual res[] for the whole speech frame
* - Find the weighted input speech wsp[] for the whole speech frame
* - Find the 2 open-loop pitch estimate
* - Set the range for searching closed-loop pitch in 1st subframe
*/
p_A = A;
for (i_subfr = 0; i_subfr < L_FRAME; i_subfr += L_SUBFR)
{
E_LPC_a_weight(p_A, Ap, GAMMA1, M);
E_UTIL_residu(Ap, &speech[i_subfr], &wsp[i_subfr], L_SUBFR);
p_A += (M + 1);
}
E_UTIL_deemph(wsp, TILT_FAC, L_FRAME, &(st->mem_wsp_df));
/* decimation of wsp[] to search pitch in LF and to reduce complexity */
E_GAIN_lp_decim2(wsp, L_FRAME, st->mem_decim2);
/* Find open loop pitch lag for whole speech frame */
if (*mode == MODE_7k)
{
/* Find open loop pitch lag for whole speech frame */
T_op = E_GAIN_open_loop_search(wsp, PIT_MIN / OPL_DECIM,
PIT_MAX / OPL_DECIM, L_FRAME / OPL_DECIM, st->mem_T0_med,
&(st->mem_ol_gain), st->mem_hf_wsp, st->mem_hp_wsp,
st->mem_ol_wght_flg);
}
else
{
/* Find open loop pitch lag for first 1/2 frame */
T_op = E_GAIN_open_loop_search(wsp, PIT_MIN / OPL_DECIM,
PIT_MAX / OPL_DECIM, (L_FRAME / 2) / OPL_DECIM, st->mem_T0_med,
&(st->mem_ol_gain), st->mem_hf_wsp, st->mem_hp_wsp,
st->mem_ol_wght_flg);
}
if (st->mem_ol_gain > 0.6)
{
st->mem_T0_med = E_GAIN_olag_median(T_op, st->mem_ol_lag);
st->mem_ada_w = 1.0F;
}
else
{
st->mem_ada_w = st->mem_ada_w * 0.9F;
}
if (st->mem_ada_w < 0.8)
{
st->mem_ol_wght_flg = 0;
}
else
{
st->mem_ol_wght_flg = 1;
}
E_DTX_pitch_tone_detection(st->vadSt, st->mem_ol_gain);
T_op *= OPL_DECIM;
if (*mode != MODE_7k)
{
/* Find open loop pitch lag for second 1/2 frame */
T_op2 = E_GAIN_open_loop_search(wsp + ((L_FRAME / 2) / OPL_DECIM),
PIT_MIN / OPL_DECIM, PIT_MAX / OPL_DECIM, (L_FRAME / 2) / OPL_DECIM,
st->mem_T0_med, &st->mem_ol_gain, st->mem_hf_wsp, st->mem_hp_wsp,
st->mem_ol_wght_flg);
if (st->mem_ol_gain > 0.6)
{
st->mem_T0_med = E_GAIN_olag_median(T_op2, st->mem_ol_lag);
st->mem_ada_w = 1.0F;
}
else
{
st->mem_ada_w = st->mem_ada_w * 0.9F;
}
if (st->mem_ada_w < 0.8)
{
st->mem_ol_wght_flg = 0;
}
else
{
st->mem_ol_wght_flg = 1;
}
E_DTX_pitch_tone_detection(st->vadSt, st->mem_ol_gain);
T_op2 *= OPL_DECIM;
}
else
{
T_op2 = T_op;
}
/*
* DTX-CNG
*/
if(*mode == MRDTX)
{
/* Buffer isf's and energy */
E_UTIL_residu(&A[3 * (M + 1)], speech, f_exc, L_FRAME);
f_tmp = 0.0;
for(i = 0; i < L_FRAME; i++)
{
f_tmp += f_exc[i] * f_exc[i];
}
E_DTX_buffer(st->dtx_encSt, isf, f_tmp, codec_mode);
/* Quantize and code the isfs */
E_DTX_exe(st->dtx_encSt, f_exc2, &prms);
/* reset speech coder memories */
E_MAIN_reset(st, 0);
/*
* Update signal for next frame.
* -> save past of speech[] and wsp[].
*/
memcpy(st->mem_speech, &st->mem_speech[L_FRAME], (L_TOTAL - L_FRAME + 460) * sizeof(Float32));
memcpy(st->mem_wsp, &st->mem_wsp[L_FRAME / OPL_DECIM],
(PIT_MAX / OPL_DECIM) * sizeof(Float32));
return(0);
}
/*
* ACELP
*/
/* Quantize and code the isfs */
if (*mode <= MODE_7k)
{
E_LPC_isf_2s3s_quantise(isf, isfq, st->mem_isf_q, indice, 4);
E_MAIN_parm_store((Word16)indice[0], &prms);
E_MAIN_parm_store((Word16)indice[1], &prms);
E_MAIN_parm_store((Word16)indice[2], &prms);
E_MAIN_parm_store((Word16)indice[3], &prms);
E_MAIN_parm_store((Word16)indice[4], &prms);
}
else
{
E_LPC_isf_2s5s_quantise(isf, isfq, st->mem_isf_q, indice, 4);
E_MAIN_parm_store((Word16)indice[0], &prms);
E_MAIN_parm_store((Word16)indice[1], &prms);
E_MAIN_parm_store((Word16)indice[2], &prms);
E_MAIN_parm_store((Word16)indice[3], &prms);
E_MAIN_parm_store((Word16)indice[4], &prms);
E_MAIN_parm_store((Word16)indice[5], &prms);
E_MAIN_parm_store((Word16)indice[6], &prms);
}
/* Convert isfs to the cosine domain */
E_LPC_isf_isp_conversion(isfq, ispnew_q, M);
if (*mode == MODE_24k)
{
/* Check stability on isf : distance between old isf and current isf */
f_tmp = 0.0F;
f_pt_tmp = st->mem_isf;
for (i=0; i < M - 1; i++)
{
f_tmp += (isf[i] - f_pt_tmp[i]) * (isf[i] - f_pt_tmp[i]);
}
stab_fac = (Float32)(1.25F - (f_tmp / 400000.0F));
if (stab_fac > 1.0F)
{
stab_fac = 1.0F;
}
if (stab_fac < 0.0F)
{
stab_fac = 0.0F;
}
memcpy(f_pt_tmp, isf, M * sizeof(Float32));
}
if (st->mem_first_frame == 1)
{
st->mem_first_frame = 0;
memcpy(st->mem_isp_q, ispnew_q, M * sizeof(Word16));
}
/* Find the interpolated isps and convert to a[] for all subframes */
E_LPC_int_isp_find(st->mem_isp_q, ispnew_q, E_ROM_interpol_frac, s_Aq);
for (i = 0; i < (NB_SUBFR * (M + 1)); i++)
{
Aq[i] = s_Aq[i] * 0.000244140625F; /* 1/4096 */
}
/* update isp memory for the next frame */
memcpy(st->mem_isp_q, ispnew_q, M * sizeof(Word16));
/*
* Find the best interpolation for quantized ISPs
*/
p_Aq = Aq;
for (i_subfr = 0; i_subfr < L_FRAME; i_subfr += L_SUBFR)
{
E_UTIL_residu(p_Aq, &speech[i_subfr], &f_exc[i_subfr], L_SUBFR);
p_Aq += (M + 1);
}
/* Buffer isf's and energy for dtx on non-speech frame */
if(vad_flag == 0)
{
f_tmp = 0.0F;
for(i = 0; i < L_FRAME; i++)
{
f_tmp += f_exc[i] * f_exc[i];
}
E_DTX_buffer(st->dtx_encSt, isf, f_tmp, codec_mode);
}
/* range for closed loop pitch search in 1st subframe */
T0_min = T_op - 8;
if (T0_min < PIT_MIN)
{
T0_min = PIT_MIN;
}
T0_max = T0_min + 15;
if (T0_max > PIT_MAX)
{
T0_max = PIT_MAX;
T0_min = T0_max - 15;
}
/*
* Loop for every subframe in the analysis frame
* ---------------------------------------------
* To find the pitch and innovation parameters. The subframe size is
* L_SUBFR and the loop is repeated L_FRAME/L_SUBFR times.
* - compute the target signal for pitch search
* - compute impulse response of weighted synthesis filter (h1[])
* - find the closed-loop pitch parameters
* - encode the pitch dealy
* - find 2 lt prediction (with / without LP filter for lt pred)
* - find 2 pitch gains and choose the best lt prediction.
* - find target vector for codebook search
* - update the impulse response h1[] for codebook search
* - correlation between target vector and impulse response
* - codebook search and encoding
* - VQ of pitch and codebook gains
* - find voicing factor and tilt of code for next subframe.
* - update states of weighting filter
* - find excitation and synthesis speech
*/
p_A = A;
p_Aq = Aq;
for (i_subfr = 0; i_subfr < L_FRAME; i_subfr += L_SUBFR)
{
pit_flag = i_subfr;
if ((i_subfr == (2 * L_SUBFR)) & (*mode > MODE_7k))
{
pit_flag = 0;
/* range for closed loop pitch search in 3rd subframe */
T0_min = T_op2 - 8;
if (T0_min < PIT_MIN)
{
T0_min = PIT_MIN;
}
T0_max = T0_min + 15;
if (T0_max > PIT_MAX)
{
T0_max = PIT_MAX;
T0_min = T0_max - 15;
}
}
/*
*
* Find the target vector for pitch search:
* ---------------------------------------
*
* |------| res[n]
* speech[n]---| A(z) |--------
* |------| | |--------| error[n] |------|
* zero -- (-)--| 1/A(z) |-----------| W(z) |-- target
* exc |--------| |------|
*
* Instead of subtracting the zero-input response of filters from
* the weighted input speech, the above configuration is used to
* compute the target vector.
*
*/
for (i = 0; i < M; i++)
{
error[i] = (Float32)(speech[i + i_subfr - 16] - st->mem_syn[i]);
}
E_UTIL_residu(p_Aq, &speech[i_subfr], &f_exc[i_subfr], L_SUBFR);
E_UTIL_synthesis(p_Aq, &f_exc[i_subfr], error + M, L_SUBFR, error, 0);
E_LPC_a_weight(p_A, Ap, GAMMA1, M);
E_UTIL_residu(Ap, error + M, xn, L_SUBFR);
E_UTIL_deemph(xn, TILT_FAC, L_SUBFR, &(st->mem_w0));
/*
* Find target in residual domain (cn[]) for innovation search.
*/
/* first half: xn[] --> cn[] */
memset(f_code, 0, M * sizeof(Float32));
memcpy(f_code + M, xn, (L_SUBFR / 2) * sizeof(Float32));
f_tmp = 0.0F;
E_UTIL_f_preemph(f_code + M, TILT_FAC, L_SUBFR / 2, &f_tmp);
E_LPC_a_weight(p_A, Ap, GAMMA1, M);
E_UTIL_synthesis(Ap, f_code + M, f_code + M, L_SUBFR / 2, f_code, 0);
E_UTIL_residu(p_Aq, f_code + M, cn, L_SUBFR / 2);
/* second half: res[] --> cn[] (approximated and faster) */
for(i = (L_SUBFR / 2); i < L_SUBFR; i++)
{
cn[i] = f_exc[i_subfr + i];
}
/*
* Compute impulse response, h1[], of weighted synthesis filter
*/
E_LPC_a_weight(p_A, Ap, GAMMA1, M);
memset(h1, 0, L_SUBFR * sizeof(Float32));
memcpy(h1, Ap, (M + 1) * sizeof(Float32));
E_UTIL_synthesis(p_Aq, h1, h1, L_SUBFR, h1 + (M + 1), 0);
f_tmp = 0.0;
E_UTIL_deemph(h1, TILT_FAC, L_SUBFR, &f_tmp);
/*
* Closed-loop fractional pitch search
*/
/* find closed loop fractional pitch lag */
if (*mode <= MODE_9k)
{
T0 = E_GAIN_closed_loop_search(&f_exc[i_subfr], xn, h1,
T0_min, T0_max, &T0_frac,
pit_flag, PIT_MIN, PIT_FR1_8b);
/* encode pitch lag */
if (pit_flag == 0) /* if 1st/3rd subframe */
{
/*
* The pitch range for the 1st/3rd subframe is encoded with
* 8 bits and is divided as follows:
* PIT_MIN to PIT_FR1-1 resolution 1/2 (frac = 0 or 2)
* PIT_FR1 to PIT_MAX resolution 1 (frac = 0)
*/
if (T0 < PIT_FR1_8b)
{
index = (Word16)(T0 * 2 + (T0_frac >> 1) - (PIT_MIN * 2));
}
else
{
index = (Word16)(T0 - PIT_FR1_8b + ((PIT_FR1_8b - PIT_MIN) * 2));
}
E_MAIN_parm_store(index, &prms);
/* find T0_min and T0_max for subframe 2 and 4 */
T0_min = T0 - 8;
if (T0_min < PIT_MIN)
{
T0_min = PIT_MIN;
}
T0_max = T0_min + 15;
if (T0_max > PIT_MAX)
{
T0_max = PIT_MAX;
T0_min = T0_max - 15;
}
}
/*
* E_IF_encode
*
*
* Parameters:
* st I: pointer to state structure
* mode I: Speech Mode
* speech I: Input speech
* serial O: Output octet structure IF2 or 16-bit serial stream
* dtx I: use DTX
*
* Function:
* Encoding and packing one frame of speech
*
* Returns:
* number of octets
*/
int E_IF_encode(void *st, Word16 req_mode, Word16 *speech, UWord8 *serial,
Word16 dtx)
{
Word16 prms[NB_PARM_MAX];
Word32 i;
Word16 frame_type, mode, reset_flag;
WB_enc_if_state *s;
s = (WB_enc_if_state *)st;
mode = req_mode;
/* check for homing frame */
reset_flag = E_IF_homing_frame_test(speech);
if (!reset_flag)
{
for (i = 0; i < L_FRAME16k; i++) /* Delete the 2 LSBs (14-bit input) */
{
speech[i] = (Word16) (speech[i] & 0xfffC);
}
E_MAIN_encode(&mode, speech, prms, s->encoder_state, dtx);
if (mode == MRDTX)
{
s->sid_update_counter--;
if (s->prev_ft == TX_SPEECH)
{
frame_type = TX_SID_FIRST;
s->sid_update_counter = 3;
}
else
{
if ((s->sid_handover_debt > 0) && (s->sid_update_counter > 2))
{
/*
* ensure extra updates are
* properly delayed after a possible SID_FIRST
*/
frame_type = TX_SID_UPDATE;
s->sid_handover_debt--;
}
else
{
if (s->sid_update_counter == 0)
{
frame_type = TX_SID_UPDATE;
s->sid_update_counter = 8;
}
else
{
frame_type = TX_NO_DATA;
mode = MRNO_DATA;
}
}
}
}
else
{
s->sid_update_counter = 8;
frame_type = TX_SPEECH;
}
s->prev_ft = frame_type;
}
/* perform homing if homing frame was detected at encoder input */
else
{
E_MAIN_reset(s->encoder_state, 1);
E_IF_sid_sync_reset(s);
E_IF_homing_coding(prms, mode);
frame_type = TX_SPEECH;
}
#ifdef IF2
return E_IF_if2_conversion(mode, prms, serial, frame_type, req_mode);
#else
return E_IF_mms_conversion(mode, prms, serial, frame_type, req_mode);
#endif
}