/* *=================================================================== * 3GPP AMR Wideband Floating-point Speech Codec *=================================================================== */ #include #include #include "typedef.h" #include "enc_util.h" #define L_FRAME 256 /* Frame size */ #define L_SUBFR 64 /* Subframe size */ #define HP_ORDER 3 #define L_INTERPOL1 4 #define L_INTERPOL2 16 #define PIT_SHARP 27853 /* pitch sharpening factor = 0.85 Q15 */ #define F_PIT_SHARP 0.85F /* pitch sharpening factor */ #define PIT_MIN 34 /* Minimum pitch lag with resolution 1/4 */ #define UP_SAMP 4 #define DIST_ISF_MAX 120 #define DIST_ISF_THRES 60 #define GAIN_PIT_THRES 0.9F #define GAIN_PIT_MIN 0.6F extern const Float32 E_ROM_corrweight[]; extern const Float32 E_ROM_inter4_1[]; extern const Word16 E_ROM_inter4_2[]; /* * E_GAIN_clip_init * * Parameters: * mem O: memory of gain of pitch clipping algorithm * * Function: * Initialises state memory * * Returns: * void */ void E_GAIN_clip_init(Float32 mem[]) { mem[0] = DIST_ISF_MAX; mem[1] = GAIN_PIT_MIN; } /* * E_GAIN_clip_test * * Parameters: * mem I: memory of gain of pitch clipping algorithm * * Function: * Gain clipping test to avoid unstable synthesis on frame erasure * * Returns: * Test result */ Word32 E_GAIN_clip_test(Float32 mem[]) { Word32 clip; clip = 0; if ((mem[0] < DIST_ISF_THRES) && (mem[1] > GAIN_PIT_THRES)) { clip = 1; } return (clip); } /* * E_GAIN_clip_isf_test * * Parameters: * isf I: isf values (in frequency domain) * mem I/O: memory of gain of pitch clipping algorithm * * Function: * Check resonance for pitch clipping algorithm * * Returns: * void */ void E_GAIN_clip_isf_test(Float32 isf[], Float32 mem[]) { Word32 i; Float32 dist, dist_min; dist_min = isf[1] - isf[0]; for (i = 2; i < M - 1; i++) { dist = isf[i] - isf[i-1]; if (dist < dist_min) { dist_min = dist; } } dist = 0.8F * mem[0] + 0.2F * dist_min; if (dist > DIST_ISF_MAX) { dist = DIST_ISF_MAX; } mem[0] = dist; return; } /* * E_GAIN_clip_pit_test * * Parameters: * gain_pit I: gain of quantized pitch * mem I/O: memory of gain of pitch clipping algorithm * * Function: * Test quantised gain of pitch for pitch clipping algorithm * * Returns: * void */ void E_GAIN_clip_pit_test(Float32 gain_pit, Float32 mem[]) { Float32 gain; gain = 0.9F * mem[1] + 0.1F * gain_pit; if (gain < GAIN_PIT_MIN) { gain = GAIN_PIT_MIN; } mem[1] = gain; return; } /* * E_GAIN_lp_decim2 * * Parameters: * x I/O: signal to process * l I: size of filtering * mem I/O: memory (size = 3) * * Function: * Decimate a vector by 2 with 2nd order fir filter. * * Returns: * void */ void E_GAIN_lp_decim2(Float32 x[], Word32 l, Float32 *mem) { Float32 x_buf[L_FRAME + 3]; Float32 temp; Word32 i, j; /* copy initial filter states into buffer */ memcpy(x_buf, mem, 3 * sizeof(Float32)); memcpy(&x_buf[3], x, l * sizeof(Float32)); for (i = 0; i < 3; i++) { mem[i] = ((x[l - 3 + i] > 1e-10) | (x[l - 3 + i] < -1e-10)) ? x[l - 3 + i] : 0; } for (i = 0, j = 0; i < l; i += 2, j++) { temp = x_buf[i] * 0.13F; temp += x_buf[i + 1] * 0.23F; temp += x_buf[i + 2] * 0.28F; temp += x_buf[i + 3] * 0.23F; temp += x_buf[i + 4] * 0.13F; x[j] = temp; } return; } /* * E_GAIN_open_loop_search * * Parameters: * wsp I: signal (end pntr) used to compute the open loop pitch * L_min I: minimum pitch lag * L_max I: maximum pitch lag * nFrame I: length of frame to compute pitch * L_0 I: old open-loop lag * gain O: open-loop pitch-gain * hp_wsp_mem I/O: memory of the highpass filter for hp_wsp[] (lg = 9) * hp_old_wsp O: highpass wsp[] * weight_flg I: is weighting function used * * Function: * Find open loop pitch lag * * Returns: * open loop pitch lag */ Word32 E_GAIN_open_loop_search(Float32 *wsp, Word32 L_min, Word32 L_max, Word32 nFrame, Word32 L_0, Float32 *gain, Float32 *hp_wsp_mem, Float32 hp_old_wsp[], UWord8 weight_flg) { Word32 i, j, k, L = 0; Float32 o, R0, R1, R2, R0_max = -1.0e23f; const Float32 *ww, *we; Float32 *data_a, *data_b, *hp_wsp, *p, *p1; ww = &E_ROM_corrweight[198]; we = &E_ROM_corrweight[98 + L_max - L_0]; for (i = L_max; i > L_min; i--) { p = &wsp[0]; p1 = &wsp[-i]; /* Compute the correlation R0 and the energy R1. */ R0 = 0.0; for (j = 0; j < nFrame; j += 2) { R0 += p[j] * p1[j]; R0 += p[j + 1] * p1[j + 1]; } /* Weighting of the correlation function. */ R0 *= *ww--; /* Weight the neighborhood of the old lag. */ if ((L_0 > 0) & (weight_flg == 1)) { R0 *= *we--; } /* Store the values if a currest maximum has been found. */ if (R0 >= R0_max) { R0_max = R0; L = i; } } data_a = hp_wsp_mem; data_b = hp_wsp_mem + HP_ORDER; hp_wsp = hp_old_wsp + L_max; for (k = 0; k < nFrame; k++) { data_b[0] = data_b[1]; data_b[1] = data_b[2]; data_b[2] = data_b[3]; data_b[HP_ORDER] = wsp[k]; o = data_b[0] * 0.83787057505665F; o += data_b[1] * -2.50975570071058F; o += data_b[2] * 2.50975570071058F; o += data_b[3] * -0.83787057505665F; o -= data_a[0] * -2.64436711600664F; o -= data_a[1] * 2.35087386625360F; o -= data_a[2] * -0.70001156927424F; data_a[2] = data_a[1]; data_a[1] = data_a[0]; data_a[0] = o; hp_wsp[k] = o; } p = &hp_wsp[0]; p1 = &hp_wsp[-L]; R0 = 0.0F; R1 = 0.0F; R2 = 0.0F; for (j = 0; j < nFrame; j++) { R1 += p1[j] * p1[j]; R2 += p[j] * p[j]; R0 += p[j] * p1[j]; } *gain = (Float32)(R0 / (sqrt(R1 * R2) + 1e-5)); memcpy(hp_old_wsp, &hp_old_wsp[nFrame], L_max * sizeof(Float32)); return(L); } /* * E_GAIN_sort * * Parameters: * n I: number of lags * ra I/O: lags / sorted lags * * Function: * Sort open-loop lags * * Returns: * void */ static void E_GAIN_sort(Word32 n, Word32 *ra) { Word32 l, j, ir, i, rra; l = (n >> 1) + 1; ir = n; for (;;) { if (l > 1) { rra = ra[--l]; } else { rra = ra[ir]; ra[ir] = ra[1]; if (--ir == 1) { ra[1] = rra; return; } } i = l; j = l << 1; while (j <= ir) { if (j < ir && ra[j] < ra[j + 1]) { ++j; } if (rra < ra[j]) { ra[i] = ra[j]; j += (i = j); } else { j = ir + 1; } } ra[i] = rra; } } /* * E_GAIN_olag_median * * Parameters: * prev_ol_lag I: previous open-loop lag * old_ol_lag I: old open-loop lags * * Function: * Median of 5 previous open-loop lags * * Returns: * median of 5 previous open-loop lags */ Word32 E_GAIN_olag_median(Word32 prev_ol_lag, Word32 old_ol_lag[5]) { Word32 tmp[6] = {0}; Word32 i; /* Use median of 5 previous open-loop lags as old lag */ for (i = 4; i > 0; i--) { old_ol_lag[i] = old_ol_lag[i-1]; } old_ol_lag[0] = prev_ol_lag; for (i = 0; i < 5; i++) { tmp[i+1] = old_ol_lag[i]; } E_GAIN_sort(5, tmp); return tmp[3]; } /* * E_GAIN_norm_corr * * Parameters: * exc I: excitation buffer * xn I: target signal * h I: weighted synthesis filter impulse response (Q15) * t0_min I: minimum value in the searched range * t0_max I: maximum value in the searched range * corr_norm O: normalized correlation (Q15) * * Function: * Find the normalized correlation between the target vector and the * filtered past excitation (correlation between target and filtered * excitation divided by the square root of energy of filtered excitation) * Size of subframe = L_SUBFR. * * Returns: * void */ static void E_GAIN_norm_corr(Float32 exc[], Float32 xn[], Float32 h[], Word32 t_min, Word32 t_max, Float32 corr_norm[]) { Float32 excf[L_SUBFR]; /* filtered past excitation */ Float32 alp, ps, norm; Word32 t, j, k; k = - t_min; /* compute the filtered excitation for the first delay t_min */ E_UTIL_f_convolve(&exc[k], h, excf); /* loop for every possible period */ for (t = t_min; t <= t_max; t++) { /* Compute correlation between xn[] and excf[] */ ps = 0.0F; alp = 0.01F; for (j = 0; j < L_SUBFR; j++) { ps += xn[j] * excf[j]; alp += excf[j] * excf[j]; } /* Compute 1/sqrt(energie of excf[]) */ norm = (Float32)(1.0F / sqrt(alp)); /* Normalize correlation = correlation * (1/sqrt(energy)) */ corr_norm[t] = ps * norm; /* update the filtered excitation excf[] for the next iteration */ if (t != t_max) { k--; for (j = L_SUBFR - 1; j > 0; j--) { excf[j] = excf[j - 1] + exc[k] * h[j]; } excf[0] = exc[k]; } } return; } /* * E_GAIN_norm_corr_interpolate * * Parameters: * x I: input vector * frac I: fraction (-4..+3) * * Function: * Interpolating the normalized correlation * * Returns: * interpolated value */ static Float32 E_GAIN_norm_corr_interpolate(Float32 *x, Word32 frac) { Float32 s, *x1, *x2; const Float32 *c1, *c2; if (frac < 0) { frac += 4; x--; } x1 = &x[0]; x2 = &x[1]; c1 = &E_ROM_inter4_1[frac]; c2 = &E_ROM_inter4_1[4 - frac]; s = x1[0] * c1[0] + x2[0] * c2[0]; s += x1[-1] * c1[4] + x2[1] * c2[4]; s += x1[-2] * c1[8] + x2[2] * c2[8]; s += x1[-3] * c1[12] + x2[3] * c2[12]; return s; } /* * E_GAIN_closed_loop_search * * Parameters: * exc I: excitation buffer * xn I: target signal * h I: weighted synthesis filter impulse response * t0_min I: minimum value in the searched range * t0_max I: maximum value in the searched range * pit_frac O: chosen fraction * i_subfr I: flag to first subframe * t0_fr2 I: minimum value for resolution 1/2 * t0_fr1 I: minimum value for resolution 1 * * Function: * Find the closed loop pitch period with 1/4 subsample resolution. * * Returns: * chosen integer pitch lag */ Word32 E_GAIN_closed_loop_search(Float32 exc[], Float32 xn[], Float32 h[], Word32 t0_min, Word32 t0_max, Word32 *pit_frac, Word32 i_subfr, Word32 t0_fr2, Word32 t0_fr1) { Float32 corr_v[15 + 2 * L_INTERPOL1 + 1]; Float32 cor_max, max, temp; Float32 *corr; Word32 i, fraction, step; Word32 t0, t_min, t_max; /* Find interval to compute normalized correlation */ t_min = t0_min - L_INTERPOL1; t_max = t0_max + L_INTERPOL1; /* allocate memory to normalized correlation vector */ corr = &corr_v[-t_min]; /* corr[t_min..t_max] */ /* Compute normalized correlation between target and filtered excitation */ E_GAIN_norm_corr(exc, xn, h, t_min, t_max, corr); /* find integer pitch */ max = corr[t0_min]; t0 = t0_min; for(i = t0_min + 1; i <= t0_max; i++) { if( corr[i] > max) { max = corr[i]; t0 = i; } } /* If first subframe and t0 >= t0_fr1, do not search fractionnal pitch */ if((i_subfr == 0) & (t0 >= t0_fr1)) { *pit_frac = 0; return(t0); } /* * Search fractionnal pitch with 1/4 subsample resolution. * Test the fractions around t0 and choose the one which maximizes * the interpolated normalized correlation. */ step = 1; /* 1/4 subsample resolution */ fraction = -3; if (((i_subfr == 0) & (t0 >= t0_fr2)) | (t0_fr2 == PIT_MIN)) { step = 2; /* 1/2 subsample resolution */ fraction = -2; } if (t0 == t0_min) { fraction = 0; } cor_max = E_GAIN_norm_corr_interpolate(&corr[t0], fraction); for (i = (fraction + step); i <= 3; i += step) { temp = E_GAIN_norm_corr_interpolate(&corr[t0], i); if (temp > cor_max) { cor_max = temp; fraction = i; } } /* limit the fraction value in the interval [0,1,2,3] */ if (fraction < 0) { fraction += 4; t0 -= 1; } *pit_frac = fraction; return (t0); } /* * E_GAIN_adaptive_codebook_excitation * * Parameters: * exc I/O: excitation buffer * T0 I: integer pitch lag * frac I: fraction of lag * L_subfr I: subframe size * * Function: * Compute the result of Word32 term prediction with fractional * interpolation of resolution 1/4. * * Returns: * interpolated signal (adaptive codebook excitation) */ void E_GAIN_adaptive_codebook_excitation(Word16 exc[], Word16 T0, Word32 frac, Word16 L_subfr) { Word32 i, j, k, L_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; j++) { L_sum = 0L; for (i = 0, k = ((UP_SAMP - 1) - frac); i < 2 * L_INTERPOL2; i++, k += UP_SAMP) { L_sum = L_sum + (x[i] * E_ROM_inter4_2[k]); } L_sum = (L_sum + 0x2000) >> 14; exc[j] = E_UTIL_saturate(L_sum); x++; } return; } /* * E_GAIN_pitch_sharpening * * Parameters: * x I/O: impulse response (or algebraic code) * pit_lag I: pitch lag * * Function: * Performs Pitch sharpening routine for one subframe. * pitch sharpening factor is 0.85 * * Returns: * void */ void E_GAIN_pitch_sharpening(Word16 *x, Word16 pit_lag) { Word32 L_tmp, i; for (i = pit_lag; i < L_SUBFR; i++) { L_tmp = x[i] << 15; L_tmp += x[i - pit_lag] * PIT_SHARP; x[i] = (Word16)((L_tmp + 0x4000) >> 15); } return; } void E_GAIN_f_pitch_sharpening(Float32 *x, Word32 pit_lag) { Word32 i; for (i = pit_lag; i < L_SUBFR; i++) { x[i] += x[i - pit_lag] * F_PIT_SHARP; } return; } /* * E_GAIN_voice_factor * * Parameters: * exc I: pitch excitation (Q_exc) * Q_exc I: exc format * gain_pit I: gain of pitch (Q14) * code I: Fixed codebook excitation (Q9) * gain_code I: gain of code (Q0) * * * Function: * Find the voicing factor (1=voice to -1=unvoiced) * Subframe length is L_SUBFR * * Returns: * factor (-1=unvoiced to 1=voiced) (Q15) */ Word32 E_GAIN_voice_factor(Word16 exc[], Word16 Q_exc, Word16 gain_pit, Word16 code[], Word16 gain_code) { Word32 i, L_tmp, tmp, exp, ener1, exp1, ener2, exp2; ener1 = E_UTIL_dot_product12(exc, exc, L_SUBFR, &exp1) >> 16; exp1 = exp1 - (Q_exc + Q_exc); L_tmp = (gain_pit * gain_pit) << 1; exp = E_UTIL_norm_l(L_tmp); tmp = (L_tmp << exp) >> 16; ener1 = (ener1 * tmp) >> 15; exp1 = (exp1 - exp) - 10; /* 10 -> gain_pit Q14 to Q9 */ ener2 = E_UTIL_dot_product12(code, code, L_SUBFR, &exp2) >> 16; exp = E_UTIL_norm_s(gain_code); tmp = gain_code << exp; tmp = (tmp * tmp) >> 15; ener2 = (ener2 * tmp) >> 15; exp2 = exp2 - (exp + exp); i = exp1 - exp2; if (i >= 0) { ener1 = ener1 >> 1; ener2 = ener2 >> (i + 1); } else { i = 1 - i; if (i < 32) { ener1 = ener1 >> i; } else { ener1 = 0; } ener2 = ener2 >> 1; } tmp = ener1 - ener2; ener1 = (ener1 + ener2) + 1; tmp = (tmp << 15) / ener1; return (tmp); }