/* *=================================================================== * 3GPP AMR Wideband Floating-point Speech Codec *=================================================================== */ #include #include "typedef.h" #include "dec_util.h" #define L_SUBFR 64 /* Subframe size */ #define L_LTPHIST 5 #define ONE_PER_3 10923 #define ONE_PER_LTPHIST 6554 #define UP_SAMP 4 #define L_INTERPOL2 16 extern const Word16 D_ROM_inter4_2[]; extern const Word16 D_ROM_pdown_unusable[]; extern const Word16 D_ROM_pdown_usable[]; extern const Word16 D_ROM_cdown_unusable[]; extern const Word16 D_ROM_cdown_usable[]; extern const Word16 D_ROM_qua_gain6b[]; extern const Word16 D_ROM_qua_gain7b[]; /* * D_GAIN_init * * Parameters: * mem O: static memory * * Function: * Initialisation of 2nd order quantiser energy predictor. * * Returns: * void */ void D_GAIN_init(Word16 *mem) { /* 4nd order quantizer energy predictor (init to -14.0 in Q10) */ mem[0] = -14336; /* past_qua_en[0] */ mem[1] = -14336; /* past_qua_en[1] */ mem[2] = -14336; /* past_qua_en[2] */ mem[3] = -14336; /* past_qua_en[3] */ /* * mem[4] = 0; past_gain_pit * mem[5] = 0; past_gain_code * mem[6] = 0; prev_gc * mem[7 - 11] = 0; pbuf[i] * mem[12 - 16] = 0; gbuf[i] * mem[17 - 21] = 0; pbuf2[i] */ memset(&mem[4], 0, 18 * sizeof(Word16)); mem[22] = 21845; /* seed */ return; } /* * D_GAIN_median * * Parameters: * buf I: previous gains * * Function: * Median of gains * * Returns: * median of 5 previous gains */ static Word16 D_GAIN_median(Word16 x[]) { Word16 x1, x2, x3, x4, x5; Word16 tmp; x1 = x[ - 2]; x2 = x[ - 1]; x3 = x[0]; x4 = x[1]; x5 = x[2]; if(x2 < x1) { tmp = x1; x1 = x2; x2 = tmp; } if(x3 < x1) { tmp = x1; x1 = x3; x3 = tmp; } if(x4 < x1) { tmp = x1; x1 = x4; x4 = tmp; } if(x5 < x1) { x5 = x1; } if(x3 < x2) { tmp = x2; x2 = x3; x3 = tmp; } if(x4 < x2) { tmp = x2; x2 = x4; x4 = tmp; } if(x5 < x2) { x5 = x2; } if(x4 < x3) { x3 = x4; } if(x5 < x3) { x3 = x5; } return(x3); } /* * D_GAIN_decode * * Parameters: * index I: Quantization index * nbits I: number of bits (6 or 7) * code I: Innovative code vector * L_subfr I: Subframe size * gain_pit O: (Q14) Quantized pitch gain * gain_code O: (Q16) Quantized codebook gain * bfi I: Bad frame indicator * prev_bfi I: Previous BF indicator * state I: State of BFH * unusable_frame I: UF indicator * vad_hist I: number of non-speech frames * mem I/O: static memory (4 words) * * * Function: * Decoding of pitch and codebook gains * * Returns: * void */ void D_GAIN_decode(Word16 index, Word16 nbits, Word16 code[], Word16 *gain_pit, Word32 *gain_cod, Word16 bfi, Word16 prev_bfi, Word16 state, Word16 unusable_frame, Word16 vad_hist, Word16 *mem) { Word32 gcode0, qua_ener, L_tmp; const Word16 * p; Word16 *past_gain_pit, *past_gain_code, *past_qua_en, *prev_gc; Word16 *gbuf, *pbuf, *pbuf2; Word16 i, tmp, exp, frac, exp_gcode0, gcode_inov; Word16 g_code; past_qua_en = mem; past_gain_pit = mem + 4; past_gain_code = mem + 5; prev_gc = mem + 6; pbuf = mem + 7; gbuf = mem + 12; pbuf2 = mem + 17; /* * Find energy of code and compute: * * L_tmp = 1.0 / sqrt(energy of code/ L_subfr) */ L_tmp = D_UTIL_dot_product12(code, code, L_SUBFR, &exp); exp = (Word16)(exp - (18 + 6)); /* exp: -18 (code in Q9), -6 (/L_subfr) */ D_UTIL_normalised_inverse_sqrt(&L_tmp, &exp); if(exp > 3) { L_tmp <<= (exp - 3); } else { L_tmp >>= (3 - exp); } gcode_inov = (Word16)(L_tmp >>16); /* g_code_inov in Q12 */ /* * Case of erasure. */ if(bfi != 0) { tmp = D_GAIN_median(&pbuf[2]); *past_gain_pit = tmp; if(*past_gain_pit > 15565) { *past_gain_pit = 15565; /* 0.95 in Q14 */ } if(unusable_frame != 0) { *gain_pit = (Word16)((D_ROM_pdown_unusable[state] * *past_gain_pit) >> 15); } else { *gain_pit = (Word16)((D_ROM_pdown_usable[state] * *past_gain_pit) >> 15); } tmp = D_GAIN_median(&gbuf[2]); if(vad_hist > 2) { *past_gain_code = tmp; } else { if(unusable_frame != 0) { *past_gain_code = (Word16)((D_ROM_cdown_unusable[state] * tmp) >> 15); } else { *past_gain_code = (Word16)((D_ROM_cdown_usable[state] * tmp) >> 15); } } /* update table of past quantized energies */ L_tmp = past_qua_en[0] + past_qua_en[1]+ past_qua_en[2] + past_qua_en[3]; qua_ener = L_tmp >> 2; qua_ener = qua_ener - 3072; /* -3 in Q10 */ if(qua_ener < - 14336) { qua_ener = -14336; /* -14 in Q10 */ } past_qua_en[3] = past_qua_en[2]; past_qua_en[2] = past_qua_en[1]; past_qua_en[1] = past_qua_en[0]; past_qua_en[0] = (Word16)qua_ener; for(i = 1; i < 5; i++) { gbuf[i - 1] = gbuf[i]; } gbuf[4] = *past_gain_code; for(i = 1; i < 5; i++) { pbuf[i - 1] = pbuf[i]; } pbuf[4] = *past_gain_pit; /* adjust gain according to energy of code */ /* past_gain_code(Q3) * gcode_inov(Q12) => Q16 */ *gain_cod = (*past_gain_code * gcode_inov) << 1; return; } /* * Compute gcode0. * = Sum(i=0,1) pred[i]*past_qua_en[i] + mean_ener - ener_code */ /* MEAN_ENER in Q24 = 0x1e000000 */ /* MA prediction coeff = {0.5, 0.4, 0.3, 0.2} in Q13 */ L_tmp = 0xF000000 + (4096 * past_qua_en[0]); /* Q13*Q10 -> Q24 */ L_tmp = L_tmp + (3277 * past_qua_en[1]); /* Q13*Q10 -> Q24 */ L_tmp = L_tmp + (2458 * past_qua_en[2]); /* Q13*Q10 -> Q24 */ L_tmp = L_tmp + (1638 * past_qua_en[3]); /* Q13*Q10 -> Q24 */ gcode0 = L_tmp >> 15; /* From Q24 to Q8 */ /* * gcode0 = pow(10.0, gcode0/20) * = pow(2, 3.321928*gcode0/20) * = pow(2, 0.166096*gcode0) */ L_tmp = (gcode0 * 5443) >> 7; /* *0.166096 in Q15 -> Q24, From Q24 to Q16 */ D_UTIL_l_extract(L_tmp, &exp_gcode0, &frac); /* Extract exponant of gcode0 */ gcode0 = D_UTIL_pow2(14, frac); /* Put 14 as exponant so that */ /* * output of Pow2() will be: * 16384 < Pow2() <= 32767 */ exp_gcode0 = (Word16)(exp_gcode0 - 14); /* Read the quantized gains */ if(nbits == 6) { p = &D_ROM_qua_gain6b[(index << 1)]; } else { p = &D_ROM_qua_gain7b[(index << 1)]; } *gain_pit = *p++; /* selected pitch gain in Q14 */ g_code = *p++; /* selected code gain in Q11 */ L_tmp = g_code * gcode0; exp_gcode0 += 5; if(exp_gcode0 >= 0) { *gain_cod = L_tmp << exp_gcode0; /* gain of code in Q16 */ } else { *gain_cod = L_tmp >> -exp_gcode0; /* gain of code in Q16 */ } if(prev_bfi == 1) { L_tmp = (*prev_gc * 5120) << 1; /* prev_gc(Q3) * 1.25(Q12) = Q16 */ /* if((*gain_cod > ((*prev_gc) * 1.25)) && (*gain_cod > 100.0)) */ if((*gain_cod > L_tmp) & (*gain_cod > 6553600)) { *gain_cod = L_tmp; } } /* keep past gain code in Q3 for frame erasure (can saturate) */ L_tmp = (*gain_cod + 0x1000) >> 13; if(L_tmp < 32768) { *past_gain_code = (Word16)L_tmp; } else { *past_gain_code = 32767; } *past_gain_pit = *gain_pit; *prev_gc = *past_gain_code; for(i = 1; i < 5; i++) { gbuf[i - 1] = gbuf[i]; } gbuf[4] = *past_gain_code; for(i = 1; i < 5; i++) { pbuf[i - 1] = pbuf[i]; } pbuf[4] = *past_gain_pit; for(i = 1; i < 5; i++) { pbuf2[i - 1] = pbuf2[i]; } pbuf2[4] = *past_gain_pit; /* adjust gain according to energy of code */ D_UTIL_l_extract(*gain_cod, &exp, &frac); L_tmp = D_UTIL_mpy_32_16(exp, frac, gcode_inov); if(L_tmp < 0xFFFFFFF) { *gain_cod = (L_tmp << 3); /* gcode_inov in Q12 */ } else { *gain_cod = 0x7FFFFFFF; } /* * qua_ener = 20*log10(g_code) * = 6.0206*log2(g_code) * = 6.0206*(log2(g_codeQ11) - 11) */ L_tmp = (Word32)(g_code); D_UTIL_log2(L_tmp, &exp, &frac); exp = (Word16)(exp - 11); L_tmp = D_UTIL_mpy_32_16(exp, frac, 24660); /* x 6.0206 in Q12 */ qua_ener = L_tmp >>3; /* result in Q10 */ /* update table of past quantized energies */ past_qua_en[3] = past_qua_en[2]; past_qua_en[2] = past_qua_en[1]; past_qua_en[1] = past_qua_en[0]; past_qua_en[0] = (Word16)qua_ener; return; } /* * D_GAIN_adaptive_control * * Parameters: * sig_in I: postfilter input signal * sig_out I/O: postfilter output signal * l_trm I: subframe size * * Function: * Adaptive gain control is used to compensate for * the gain difference between the non-emphasized excitation and * emphasized excitation. * * Returns: * void */ void D_GAIN_adaptive_control(Word16 *sig_in, Word16 *sig_out, Word16 l_trm) { Word32 s, temp, i, exp; Word32 gain_in, gain_out, g0; /* calculate gain_out with exponent */ temp = sig_out[0] >> 2; s = temp * temp; for(i = 1; i < l_trm; i++) { temp = sig_out[i] >> 2; s += temp * temp; } s <<= 1; if(s == 0) { return; } exp = (D_UTIL_norm_l(s) - 1); if(exp >= 0) { gain_out = ((s << exp) + 0x8000) >> 16; } else { gain_out = ((s >> -exp) + 0x8000) >> 16; } /* calculate gain_in with exponent */ temp = sig_in[0] >> 2; s = temp * temp; for(i = 1; i < l_trm; i++) { temp = sig_in[i] >> 2; s += temp * temp; } s <<= 1; if(s == 0) { g0 = 0; } else { i = D_UTIL_norm_l(s); s = ((s << i) + 0x8000) >> 16; if((s < 32768) & (s > 0)) { gain_in = s; } else { gain_in = 32767; } exp = exp - i; /* * g0 = sqrt(gain_in/gain_out) */ s = (gain_out << 15) / gain_in; s = s << (7 - exp); /* s = gain_out / gain_in */ s = D_UTIL_inverse_sqrt(s); g0 = ((s << 9) + 0x8000) >> 16; } /* sig_out(n) = gain(n) sig_out(n) */ for(i = 0; i < l_trm; i++) { s = (sig_out[i] * g0) >> 13; sig_out[i] = D_UTIL_saturate(s); } return; } /* * D_GAIN_insert_lag * * Parameters: * array I/O: pitch lag history * n I: history size * x I: lag value * * Function: * Insert lag into correct location * * Returns: * void */ static void D_GAIN_insert_lag(Word16 array[], Word32 n, Word16 x) { Word32 i; for(i = n - 1; i >= 0; i--) { if(x < array[i]) { array[i + 1] = array[i]; } else { break; } } array[i + 1] = x; } /* * D_GAIN_sort_lag * * Parameters: * array I/O: pitch lag history * n I: history size * * Function: * Sorting of the lag history * * Returns: * void */ static void D_GAIN_sort_lag(Word16 array[], Word16 n) { Word32 i; for(i = 0; i < n; i++) { D_GAIN_insert_lag(array, i, array[i]); } } /* * D_GAIN_lag_concealment_init * * Parameters: * lag_hist O: pitch lag history * * Function: * Initialise lag history to 64 * * Returns: * void */ void D_GAIN_lag_concealment_init(Word16 lag_hist[]) { Word32 i; for(i = 0; i < L_LTPHIST; i++) { lag_hist[i] = 64; } } /* * D_GAIN_lag_concealment * * Parameters: * gain_hist I: gain history * lag_hist I: pitch lag history * T0 O: current lag * old_T0 I: previous lag * seed I/O: seed for random * unusable_frame I: lost frame * * Function: * Concealment of LTP lags during bad frames * * Returns: * void */ void D_GAIN_lag_concealment(Word16 gain_hist[], Word16 lag_hist[], Word32 *T0, Word16 *old_T0, Word16 *seed, Word16 unusable_frame) { Word32 i, lagDif, tmp, tmp2, D2, meanLag = 0; Word16 lag_hist2[L_LTPHIST] = {0}; Word16 maxLag, minLag, lastLag; Word16 minGain, lastGain, secLastGain; Word16 D; /* * Is lag index such that it can be aplied directly * or does it has to be subtituted */ lastGain = gain_hist[4]; secLastGain = gain_hist[3]; lastLag = lag_hist[0]; /* SMALLEST history lag */ minLag = lag_hist[0]; for(i = 1; i < L_LTPHIST; i++) { if(lag_hist[i] < minLag) { minLag = lag_hist[i]; } } /* BIGGEST history lag */ maxLag = lag_hist[0]; for(i = 1; i < L_LTPHIST; i++) { if(lag_hist[i] > maxLag) { maxLag = lag_hist[i]; } } /* SMALLEST history gain */ minGain = gain_hist[0]; for(i = 1; i < L_LTPHIST; i++) { if(gain_hist[i] < minGain) { minGain = gain_hist[i]; } } /* Difference between MAX and MIN lag */ lagDif = maxLag - minLag; if(unusable_frame != 0) { /* * LTP-lag for RX_SPEECH_LOST * Recognition of the LTP-history */ if((minGain > 8192) & (lagDif < 10)) { *T0 = *old_T0; } else if((lastGain > 8192) && (secLastGain > 8192)) { *T0 = lag_hist[0]; } else { /* * SORT * The sorting of the lag history */ for(i = 0; i < L_LTPHIST; i++) { lag_hist2[i] = lag_hist[i]; } D_GAIN_sort_lag(lag_hist2, 5); /* * Lag is weighted towards bigger lags * and random variation is added */ lagDif = (lag_hist2[4] - lag_hist2[2]); if(lagDif > 40) { lagDif = 40; } D = D_UTIL_random(seed); /* D={-1, ...,1} */ /* D2={-lagDif/2..lagDif/2} */ tmp = lagDif >> 1; D2 = (tmp * D) >> 15; tmp = (lag_hist2[2] + lag_hist2[3]) + lag_hist2[4]; *T0 = ((tmp * ONE_PER_3) >> 15) + D2; } /* New lag is not allowed to be bigger or smaller than last lag values */ if(*T0 > maxLag) { *T0 = maxLag; } if(*T0 < minLag) { *T0 = minLag; } } else { /* * LTP-lag for RX_BAD_FRAME * MEAN lag */ meanLag = 0; for(i = 0; i < L_LTPHIST; i++) { meanLag = meanLag + lag_hist[i]; } meanLag = (meanLag * ONE_PER_LTPHIST) >> 15; tmp = *T0 - maxLag; tmp2 = *T0 - lastLag; if((lagDif < 10) & (*T0 > (minLag - 5)) & (tmp < 5)) { *T0 = *T0; } else if((lastGain > 8192) & (secLastGain > 8192) & ((tmp2 > - 10) & (tmp2 < 10))) { *T0 = *T0; } else if((minGain < 6554) & (lastGain == minGain) & ((*T0 > minLag) & (*T0 < maxLag))) { *T0 = *T0; } else if((lagDif < 70) & (*T0 > minLag) & (*T0 < maxLag)) { *T0 = *T0; } else if((*T0 > meanLag) & (*T0 < maxLag)) { *T0 = *T0; } else { if((minGain > 8192) & (lagDif < 10)) { *T0 = lag_hist[0]; } else if((lastGain > 8192) & (secLastGain > 8192)) { *T0 = lag_hist[0]; } else { /* * SORT * The sorting of the lag history */ for(i = 0; i < L_LTPHIST; i++) { lag_hist2[i] = lag_hist[i]; } D_GAIN_sort_lag(lag_hist2, 5); /* * Lag is weighted towards bigger lags * and random variation is added */ lagDif = lag_hist2[4] - lag_hist2[2]; if(lagDif > 40) { lagDif = 40; } D = D_UTIL_random(seed); /* D={-1,.., 1} */ /* D2={-lagDif/2..lagDif/2} */ tmp = lagDif >> 1; D2 = (tmp * D) >> 15; tmp = (lag_hist2[2] + lag_hist2[3]) + lag_hist2[4]; *T0 = ((tmp * ONE_PER_3) >> 15) + D2; } /* * New lag is not allowed to be bigger or * smaller than last lag values */ if(*T0 > maxLag) { *T0 = maxLag; } if(*T0 < minLag) { *T0 = minLag; } } } } /* * D_GAIN_adaptive_codebook_excitation * * Parameters: * exc I/O: excitation buffer * T0 I: integer pitch lag * frac I: fraction of lag * * Function: * Compute the result of Word32 term prediction with fractional * interpolation of resolution 1/4. * * Returns: * interpolated signal (adaptive codebook excitation) */ void D_GAIN_adaptive_codebook_excitation(Word16 exc[], Word32 T0, Word32 frac) { Word32 i, j, k, sum; Word16 *x; x = &exc[ - T0]; frac = -(frac); if(frac < 0) { frac = (frac + UP_SAMP); x--; } x = x - L_INTERPOL2 + 1; for(j = 0; j < L_SUBFR + 1; j++) { sum = 0L; for(i = 0, k = ((UP_SAMP - 1) - frac); i < 2 * L_INTERPOL2; i++, k += UP_SAMP) { sum += x[i] * D_ROM_inter4_2[k]; } sum = (sum + 0x2000) >> 14; exc[j] = D_UTIL_saturate(sum); x++; } return; } /* * D_GAIN_pitch_sharpening * * Parameters: * x I/O: impulse response (or algebraic code) * pit_lag I: pitch lag * sharp I: (Q15) pitch sharpening factor * * Function: * Performs Pitch sharpening routine for one subframe. * * Returns: * void */ void D_GAIN_pitch_sharpening(Word16 *x, Word32 pit_lag, Word16 sharp) { Word32 i; Word32 tmp; for(i = pit_lag; i < L_SUBFR; i++) { tmp = x[i] << 15; tmp += x[i - pit_lag] * sharp; x[i] = (Word16)((tmp + 0x4000) >> 15); } return; } /* * D_GAIN_find_voice_factor * * Parameters: * exc I: pitch excitation * Q_exc I: exc format * gain_pit I: (Q14) gain of pitch * code I: (Q9) fixed codebook excitation * gain_code I: (Q0) gain of code * L_subfr I: subframe length * * Function: * Find the voicing factor. * * Returns: * (Q15) 1=voice to -1=unvoiced */ Word16 D_GAIN_find_voice_factor(Word16 exc[], Word16 Q_exc, Word16 gain_pit, Word16 code[], Word16 gain_code, Word16 L_subfr) { Word32 tmp, ener1, ener2, i; Word16 exp, exp1, exp2; ener1 = (D_UTIL_dot_product12(exc, exc, L_subfr, &exp1)) >> 16; exp1 = (Word16)(exp1 - (Q_exc + Q_exc)); tmp = (gain_pit * gain_pit) << 1; exp = D_UTIL_norm_l(tmp); tmp = (tmp << exp) >> 16; ener1 = (ener1 * tmp) >> 15; exp1 = (Word16)((exp1 - exp) - 10); /* 10 -> gain_pit Q14 to Q9 */ ener2 = D_UTIL_dot_product12(code, code, L_subfr, &exp2) >> 16; exp = D_UTIL_norm_s(gain_code); tmp = gain_code << exp; tmp = (tmp * tmp) >> 15; ener2 = (ener2 * tmp) >> 15; exp2 = (Word16)(exp2 - (exp << 1)); i = exp1 - exp2; if(i >= 0) { ener1 = ener1 >> 1; ener2 = ener2 >> (i + 1); } else if(i > (-16)) { ener1 = ener1 >> (1 - i); ener2 = ener2 >> 1; } else { ener1 = 0; ener2 = ener2 >> 1; } tmp = ener1 - ener2; ener1 = (ener1 + ener2) + 1; tmp = (tmp << 15) / ener1; return((Word16)tmp); }