yading@10: /* yading@10: * QCELP decoder yading@10: * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: /** yading@10: * @file yading@10: * QCELP decoder yading@10: * @author Reynaldo H. Verdejo Pinochet yading@10: * @remark FFmpeg merging spearheaded by Kenan Gillet yading@10: * @remark Development mentored by Benjamin Larson yading@10: */ yading@10: yading@10: #include yading@10: yading@10: #include "libavutil/channel_layout.h" yading@10: #include "libavutil/float_dsp.h" yading@10: #include "avcodec.h" yading@10: #include "internal.h" yading@10: #include "get_bits.h" yading@10: #include "qcelpdata.h" yading@10: #include "celp_filters.h" yading@10: #include "acelp_filters.h" yading@10: #include "acelp_vectors.h" yading@10: #include "lsp.h" yading@10: yading@10: #undef NDEBUG yading@10: #include yading@10: yading@10: typedef enum { yading@10: I_F_Q = -1, /**< insufficient frame quality */ yading@10: SILENCE, yading@10: RATE_OCTAVE, yading@10: RATE_QUARTER, yading@10: RATE_HALF, yading@10: RATE_FULL yading@10: } qcelp_packet_rate; yading@10: yading@10: typedef struct { yading@10: GetBitContext gb; yading@10: qcelp_packet_rate bitrate; yading@10: QCELPFrame frame; /**< unpacked data frame */ yading@10: yading@10: uint8_t erasure_count; yading@10: uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */ yading@10: float prev_lspf[10]; yading@10: float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */ yading@10: float pitch_synthesis_filter_mem[303]; yading@10: float pitch_pre_filter_mem[303]; yading@10: float rnd_fir_filter_mem[180]; yading@10: float formant_mem[170]; yading@10: float last_codebook_gain; yading@10: int prev_g1[2]; yading@10: int prev_bitrate; yading@10: float pitch_gain[4]; yading@10: uint8_t pitch_lag[4]; yading@10: uint16_t first16bits; yading@10: uint8_t warned_buf_mismatch_bitrate; yading@10: yading@10: /* postfilter */ yading@10: float postfilter_synth_mem[10]; yading@10: float postfilter_agc_mem; yading@10: float postfilter_tilt_mem; yading@10: } QCELPContext; yading@10: yading@10: /** yading@10: * Initialize the speech codec according to the specification. yading@10: * yading@10: * TIA/EIA/IS-733 2.4.9 yading@10: */ yading@10: static av_cold int qcelp_decode_init(AVCodecContext *avctx) yading@10: { yading@10: QCELPContext *q = avctx->priv_data; yading@10: int i; yading@10: yading@10: avctx->channels = 1; yading@10: avctx->channel_layout = AV_CH_LAYOUT_MONO; yading@10: avctx->sample_fmt = AV_SAMPLE_FMT_FLT; yading@10: yading@10: for (i = 0; i < 10; i++) yading@10: q->prev_lspf[i] = (i + 1) / 11.; yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Decode the 10 quantized LSP frequencies from the LSPV/LSP yading@10: * transmission codes of any bitrate and check for badly received packets. yading@10: * yading@10: * @param q the context yading@10: * @param lspf line spectral pair frequencies yading@10: * yading@10: * @return 0 on success, -1 if the packet is badly received yading@10: * yading@10: * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3 yading@10: */ yading@10: static int decode_lspf(QCELPContext *q, float *lspf) yading@10: { yading@10: int i; yading@10: float tmp_lspf, smooth, erasure_coeff; yading@10: const float *predictors; yading@10: yading@10: if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) { yading@10: predictors = q->prev_bitrate != RATE_OCTAVE && yading@10: q->prev_bitrate != I_F_Q ? q->prev_lspf yading@10: : q->predictor_lspf; yading@10: yading@10: if (q->bitrate == RATE_OCTAVE) { yading@10: q->octave_count++; yading@10: yading@10: for (i = 0; i < 10; i++) { yading@10: q->predictor_lspf[i] = yading@10: lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR yading@10: : -QCELP_LSP_SPREAD_FACTOR) + yading@10: predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR + yading@10: (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11); yading@10: } yading@10: smooth = q->octave_count < 10 ? .875 : 0.1; yading@10: } else { yading@10: erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR; yading@10: yading@10: assert(q->bitrate == I_F_Q); yading@10: yading@10: if (q->erasure_count > 1) yading@10: erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7; yading@10: yading@10: for (i = 0; i < 10; i++) { yading@10: q->predictor_lspf[i] = yading@10: lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 + yading@10: erasure_coeff * predictors[i]; yading@10: } yading@10: smooth = 0.125; yading@10: } yading@10: yading@10: // Check the stability of the LSP frequencies. yading@10: lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR); yading@10: for (i = 1; i < 10; i++) yading@10: lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR); yading@10: yading@10: lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR); yading@10: for (i = 9; i > 0; i--) yading@10: lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR); yading@10: yading@10: // Low-pass filter the LSP frequencies. yading@10: ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10); yading@10: } else { yading@10: q->octave_count = 0; yading@10: yading@10: tmp_lspf = 0.; yading@10: for (i = 0; i < 5; i++) { yading@10: lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001; yading@10: lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001; yading@10: } yading@10: yading@10: // Check for badly received packets. yading@10: if (q->bitrate == RATE_QUARTER) { yading@10: if (lspf[9] <= .70 || lspf[9] >= .97) yading@10: return -1; yading@10: for (i = 3; i < 10; i++) yading@10: if (fabs(lspf[i] - lspf[i - 2]) < .08) yading@10: return -1; yading@10: } else { yading@10: if (lspf[9] <= .66 || lspf[9] >= .985) yading@10: return -1; yading@10: for (i = 4; i < 10; i++) yading@10: if (fabs(lspf[i] - lspf[i - 4]) < .0931) yading@10: return -1; yading@10: } yading@10: } yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Convert codebook transmission codes to GAIN and INDEX. yading@10: * yading@10: * @param q the context yading@10: * @param gain array holding the decoded gain yading@10: * yading@10: * TIA/EIA/IS-733 2.4.6.2 yading@10: */ yading@10: static void decode_gain_and_index(QCELPContext *q, float *gain) yading@10: { yading@10: int i, subframes_count, g1[16]; yading@10: float slope; yading@10: yading@10: if (q->bitrate >= RATE_QUARTER) { yading@10: switch (q->bitrate) { yading@10: case RATE_FULL: subframes_count = 16; break; yading@10: case RATE_HALF: subframes_count = 4; break; yading@10: default: subframes_count = 5; yading@10: } yading@10: for (i = 0; i < subframes_count; i++) { yading@10: g1[i] = 4 * q->frame.cbgain[i]; yading@10: if (q->bitrate == RATE_FULL && !((i + 1) & 3)) { yading@10: g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32); yading@10: } yading@10: yading@10: gain[i] = qcelp_g12ga[g1[i]]; yading@10: yading@10: if (q->frame.cbsign[i]) { yading@10: gain[i] = -gain[i]; yading@10: q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127; yading@10: } yading@10: } yading@10: yading@10: q->prev_g1[0] = g1[i - 2]; yading@10: q->prev_g1[1] = g1[i - 1]; yading@10: q->last_codebook_gain = qcelp_g12ga[g1[i - 1]]; yading@10: yading@10: if (q->bitrate == RATE_QUARTER) { yading@10: // Provide smoothing of the unvoiced excitation energy. yading@10: gain[7] = gain[4]; yading@10: gain[6] = 0.4 * gain[3] + 0.6 * gain[4]; yading@10: gain[5] = gain[3]; yading@10: gain[4] = 0.8 * gain[2] + 0.2 * gain[3]; yading@10: gain[3] = 0.2 * gain[1] + 0.8 * gain[2]; yading@10: gain[2] = gain[1]; yading@10: gain[1] = 0.6 * gain[0] + 0.4 * gain[1]; yading@10: } yading@10: } else if (q->bitrate != SILENCE) { yading@10: if (q->bitrate == RATE_OCTAVE) { yading@10: g1[0] = 2 * q->frame.cbgain[0] + yading@10: av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54); yading@10: subframes_count = 8; yading@10: } else { yading@10: assert(q->bitrate == I_F_Q); yading@10: yading@10: g1[0] = q->prev_g1[1]; yading@10: switch (q->erasure_count) { yading@10: case 1 : break; yading@10: case 2 : g1[0] -= 1; break; yading@10: case 3 : g1[0] -= 2; break; yading@10: default: g1[0] -= 6; yading@10: } yading@10: if (g1[0] < 0) yading@10: g1[0] = 0; yading@10: subframes_count = 4; yading@10: } yading@10: // This interpolation is done to produce smoother background noise. yading@10: slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count; yading@10: for (i = 1; i <= subframes_count; i++) yading@10: gain[i - 1] = q->last_codebook_gain + slope * i; yading@10: yading@10: q->last_codebook_gain = gain[i - 2]; yading@10: q->prev_g1[0] = q->prev_g1[1]; yading@10: q->prev_g1[1] = g1[0]; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * If the received packet is Rate 1/4 a further sanity check is made of the yading@10: * codebook gain. yading@10: * yading@10: * @param cbgain the unpacked cbgain array yading@10: * @return -1 if the sanity check fails, 0 otherwise yading@10: * yading@10: * TIA/EIA/IS-733 2.4.8.7.3 yading@10: */ yading@10: static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain) yading@10: { yading@10: int i, diff, prev_diff = 0; yading@10: yading@10: for (i = 1; i < 5; i++) { yading@10: diff = cbgain[i] - cbgain[i-1]; yading@10: if (FFABS(diff) > 10) yading@10: return -1; yading@10: else if (FFABS(diff - prev_diff) > 12) yading@10: return -1; yading@10: prev_diff = diff; yading@10: } yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Compute the scaled codebook vector Cdn From INDEX and GAIN yading@10: * for all rates. yading@10: * yading@10: * The specification lacks some information here. yading@10: * yading@10: * TIA/EIA/IS-733 has an omission on the codebook index determination yading@10: * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says yading@10: * you have to subtract the decoded index parameter from the given scaled yading@10: * codebook vector index 'n' to get the desired circular codebook index, but yading@10: * it does not mention that you have to clamp 'n' to [0-9] in order to get yading@10: * RI-compliant results. yading@10: * yading@10: * The reason for this mistake seems to be the fact they forgot to mention you yading@10: * have to do these calculations per codebook subframe and adjust given yading@10: * equation values accordingly. yading@10: * yading@10: * @param q the context yading@10: * @param gain array holding the 4 pitch subframe gain values yading@10: * @param cdn_vector array for the generated scaled codebook vector yading@10: */ yading@10: static void compute_svector(QCELPContext *q, const float *gain, yading@10: float *cdn_vector) yading@10: { yading@10: int i, j, k; yading@10: uint16_t cbseed, cindex; yading@10: float *rnd, tmp_gain, fir_filter_value; yading@10: yading@10: switch (q->bitrate) { yading@10: case RATE_FULL: yading@10: for (i = 0; i < 16; i++) { yading@10: tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO; yading@10: cindex = -q->frame.cindex[i]; yading@10: for (j = 0; j < 10; j++) yading@10: *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127]; yading@10: } yading@10: break; yading@10: case RATE_HALF: yading@10: for (i = 0; i < 4; i++) { yading@10: tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO; yading@10: cindex = -q->frame.cindex[i]; yading@10: for (j = 0; j < 40; j++) yading@10: *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127]; yading@10: } yading@10: break; yading@10: case RATE_QUARTER: yading@10: cbseed = (0x0003 & q->frame.lspv[4]) << 14 | yading@10: (0x003F & q->frame.lspv[3]) << 8 | yading@10: (0x0060 & q->frame.lspv[2]) << 1 | yading@10: (0x0007 & q->frame.lspv[1]) << 3 | yading@10: (0x0038 & q->frame.lspv[0]) >> 3; yading@10: rnd = q->rnd_fir_filter_mem + 20; yading@10: for (i = 0; i < 8; i++) { yading@10: tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0); yading@10: for (k = 0; k < 20; k++) { yading@10: cbseed = 521 * cbseed + 259; yading@10: *rnd = (int16_t) cbseed; yading@10: yading@10: // FIR filter yading@10: fir_filter_value = 0.0; yading@10: for (j = 0; j < 10; j++) yading@10: fir_filter_value += qcelp_rnd_fir_coefs[j] * yading@10: (rnd[-j] + rnd[-20+j]); yading@10: yading@10: fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10]; yading@10: *cdn_vector++ = tmp_gain * fir_filter_value; yading@10: rnd++; yading@10: } yading@10: } yading@10: memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, yading@10: 20 * sizeof(float)); yading@10: break; yading@10: case RATE_OCTAVE: yading@10: cbseed = q->first16bits; yading@10: for (i = 0; i < 8; i++) { yading@10: tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0); yading@10: for (j = 0; j < 20; j++) { yading@10: cbseed = 521 * cbseed + 259; yading@10: *cdn_vector++ = tmp_gain * (int16_t) cbseed; yading@10: } yading@10: } yading@10: break; yading@10: case I_F_Q: yading@10: cbseed = -44; // random codebook index yading@10: for (i = 0; i < 4; i++) { yading@10: tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO; yading@10: for (j = 0; j < 40; j++) yading@10: *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127]; yading@10: } yading@10: break; yading@10: case SILENCE: yading@10: memset(cdn_vector, 0, 160 * sizeof(float)); yading@10: break; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Apply generic gain control. yading@10: * yading@10: * @param v_out output vector yading@10: * @param v_in gain-controlled vector yading@10: * @param v_ref vector to control gain of yading@10: * yading@10: * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6 yading@10: */ yading@10: static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in) yading@10: { yading@10: int i; yading@10: yading@10: for (i = 0; i < 160; i += 40) { yading@10: float res = avpriv_scalarproduct_float_c(v_ref + i, v_ref + i, 40); yading@10: ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, res, 40); yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Apply filter in pitch-subframe steps. yading@10: * yading@10: * @param memory buffer for the previous state of the filter yading@10: * - must be able to contain 303 elements yading@10: * - the 143 first elements are from the previous state yading@10: * - the next 160 are for output yading@10: * @param v_in input filter vector yading@10: * @param gain per-subframe gain array, each element is between 0.0 and 2.0 yading@10: * @param lag per-subframe lag array, each element is yading@10: * - between 16 and 143 if its corresponding pfrac is 0, yading@10: * - between 16 and 139 otherwise yading@10: * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0 yading@10: * otherwise yading@10: * yading@10: * @return filter output vector yading@10: */ yading@10: static const float *do_pitchfilter(float memory[303], const float v_in[160], yading@10: const float gain[4], const uint8_t *lag, yading@10: const uint8_t pfrac[4]) yading@10: { yading@10: int i, j; yading@10: float *v_lag, *v_out; yading@10: const float *v_len; yading@10: yading@10: v_out = memory + 143; // Output vector starts at memory[143]. yading@10: yading@10: for (i = 0; i < 4; i++) { yading@10: if (gain[i]) { yading@10: v_lag = memory + 143 + 40 * i - lag[i]; yading@10: for (v_len = v_in + 40; v_in < v_len; v_in++) { yading@10: if (pfrac[i]) { // If it is a fractional lag... yading@10: for (j = 0, *v_out = 0.; j < 4; j++) yading@10: *v_out += qcelp_hammsinc_table[j] * (v_lag[j - 4] + v_lag[3 - j]); yading@10: } else yading@10: *v_out = *v_lag; yading@10: yading@10: *v_out = *v_in + gain[i] * *v_out; yading@10: yading@10: v_lag++; yading@10: v_out++; yading@10: } yading@10: } else { yading@10: memcpy(v_out, v_in, 40 * sizeof(float)); yading@10: v_in += 40; yading@10: v_out += 40; yading@10: } yading@10: } yading@10: yading@10: memmove(memory, memory + 160, 143 * sizeof(float)); yading@10: return memory + 143; yading@10: } yading@10: yading@10: /** yading@10: * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector. yading@10: * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2 yading@10: * yading@10: * @param q the context yading@10: * @param cdn_vector the scaled codebook vector yading@10: */ yading@10: static void apply_pitch_filters(QCELPContext *q, float *cdn_vector) yading@10: { yading@10: int i; yading@10: const float *v_synthesis_filtered, *v_pre_filtered; yading@10: yading@10: if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE || yading@10: (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) { yading@10: yading@10: if (q->bitrate >= RATE_HALF) { yading@10: // Compute gain & lag for the whole frame. yading@10: for (i = 0; i < 4; i++) { yading@10: q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0; yading@10: yading@10: q->pitch_lag[i] = q->frame.plag[i] + 16; yading@10: } yading@10: } else { yading@10: float max_pitch_gain; yading@10: yading@10: if (q->bitrate == I_F_Q) { yading@10: if (q->erasure_count < 3) yading@10: max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1); yading@10: else yading@10: max_pitch_gain = 0.0; yading@10: } else { yading@10: assert(q->bitrate == SILENCE); yading@10: max_pitch_gain = 1.0; yading@10: } yading@10: for (i = 0; i < 4; i++) yading@10: q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain); yading@10: yading@10: memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac)); yading@10: } yading@10: yading@10: // pitch synthesis filter yading@10: v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem, yading@10: cdn_vector, q->pitch_gain, yading@10: q->pitch_lag, q->frame.pfrac); yading@10: yading@10: // pitch prefilter update yading@10: for (i = 0; i < 4; i++) yading@10: q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0); yading@10: yading@10: v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem, yading@10: v_synthesis_filtered, yading@10: q->pitch_gain, q->pitch_lag, yading@10: q->frame.pfrac); yading@10: yading@10: apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered); yading@10: } else { yading@10: memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17, 143 * sizeof(float)); yading@10: memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float)); yading@10: memset(q->pitch_gain, 0, sizeof(q->pitch_gain)); yading@10: memset(q->pitch_lag, 0, sizeof(q->pitch_lag)); yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Reconstruct LPC coefficients from the line spectral pair frequencies yading@10: * and perform bandwidth expansion. yading@10: * yading@10: * @param lspf line spectral pair frequencies yading@10: * @param lpc linear predictive coding coefficients yading@10: * yading@10: * @note: bandwidth_expansion_coeff could be precalculated into a table yading@10: * but it seems to be slower on x86 yading@10: * yading@10: * TIA/EIA/IS-733 2.4.3.3.5 yading@10: */ yading@10: static void lspf2lpc(const float *lspf, float *lpc) yading@10: { yading@10: double lsp[10]; yading@10: double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF; yading@10: int i; yading@10: yading@10: for (i = 0; i < 10; i++) yading@10: lsp[i] = cos(M_PI * lspf[i]); yading@10: yading@10: ff_acelp_lspd2lpc(lsp, lpc, 5); yading@10: yading@10: for (i = 0; i < 10; i++) { yading@10: lpc[i] *= bandwidth_expansion_coeff; yading@10: bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Interpolate LSP frequencies and compute LPC coefficients yading@10: * for a given bitrate & pitch subframe. yading@10: * yading@10: * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2 yading@10: * yading@10: * @param q the context yading@10: * @param curr_lspf LSP frequencies vector of the current frame yading@10: * @param lpc float vector for the resulting LPC yading@10: * @param subframe_num frame number in decoded stream yading@10: */ yading@10: static void interpolate_lpc(QCELPContext *q, const float *curr_lspf, yading@10: float *lpc, const int subframe_num) yading@10: { yading@10: float interpolated_lspf[10]; yading@10: float weight; yading@10: yading@10: if (q->bitrate >= RATE_QUARTER) yading@10: weight = 0.25 * (subframe_num + 1); yading@10: else if (q->bitrate == RATE_OCTAVE && !subframe_num) yading@10: weight = 0.625; yading@10: else yading@10: weight = 1.0; yading@10: yading@10: if (weight != 1.0) { yading@10: ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf, yading@10: weight, 1.0 - weight, 10); yading@10: lspf2lpc(interpolated_lspf, lpc); yading@10: } else if (q->bitrate >= RATE_QUARTER || yading@10: (q->bitrate == I_F_Q && !subframe_num)) yading@10: lspf2lpc(curr_lspf, lpc); yading@10: else if (q->bitrate == SILENCE && !subframe_num) yading@10: lspf2lpc(q->prev_lspf, lpc); yading@10: } yading@10: yading@10: static qcelp_packet_rate buf_size2bitrate(const int buf_size) yading@10: { yading@10: switch (buf_size) { yading@10: case 35: return RATE_FULL; yading@10: case 17: return RATE_HALF; yading@10: case 8: return RATE_QUARTER; yading@10: case 4: return RATE_OCTAVE; yading@10: case 1: return SILENCE; yading@10: } yading@10: yading@10: return I_F_Q; yading@10: } yading@10: yading@10: /** yading@10: * Determine the bitrate from the frame size and/or the first byte of the frame. yading@10: * yading@10: * @param avctx the AV codec context yading@10: * @param buf_size length of the buffer yading@10: * @param buf the bufffer yading@10: * yading@10: * @return the bitrate on success, yading@10: * I_F_Q if the bitrate cannot be satisfactorily determined yading@10: * yading@10: * TIA/EIA/IS-733 2.4.8.7.1 yading@10: */ yading@10: static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, yading@10: const int buf_size, yading@10: const uint8_t **buf) yading@10: { yading@10: qcelp_packet_rate bitrate; yading@10: yading@10: if ((bitrate = buf_size2bitrate(buf_size)) >= 0) { yading@10: if (bitrate > **buf) { yading@10: QCELPContext *q = avctx->priv_data; yading@10: if (!q->warned_buf_mismatch_bitrate) { yading@10: av_log(avctx, AV_LOG_WARNING, yading@10: "Claimed bitrate and buffer size mismatch.\n"); yading@10: q->warned_buf_mismatch_bitrate = 1; yading@10: } yading@10: bitrate = **buf; yading@10: } else if (bitrate < **buf) { yading@10: av_log(avctx, AV_LOG_ERROR, yading@10: "Buffer is too small for the claimed bitrate.\n"); yading@10: return I_F_Q; yading@10: } yading@10: (*buf)++; yading@10: } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) { yading@10: av_log(avctx, AV_LOG_WARNING, yading@10: "Bitrate byte is missing, guessing the bitrate from packet size.\n"); yading@10: } else yading@10: return I_F_Q; yading@10: yading@10: if (bitrate == SILENCE) { yading@10: // FIXME: Remove this warning when tested with samples. yading@10: avpriv_request_sample(avctx, "Blank frame handling"); yading@10: } yading@10: return bitrate; yading@10: } yading@10: yading@10: static void warn_insufficient_frame_quality(AVCodecContext *avctx, yading@10: const char *message) yading@10: { yading@10: av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", yading@10: avctx->frame_number, message); yading@10: } yading@10: yading@10: static void postfilter(QCELPContext *q, float *samples, float *lpc) yading@10: { yading@10: static const float pow_0_775[10] = { yading@10: 0.775000, 0.600625, 0.465484, 0.360750, 0.279582, yading@10: 0.216676, 0.167924, 0.130141, 0.100859, 0.078166 yading@10: }, pow_0_625[10] = { yading@10: 0.625000, 0.390625, 0.244141, 0.152588, 0.095367, yading@10: 0.059605, 0.037253, 0.023283, 0.014552, 0.009095 yading@10: }; yading@10: float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160]; yading@10: int n; yading@10: yading@10: for (n = 0; n < 10; n++) { yading@10: lpc_s[n] = lpc[n] * pow_0_625[n]; yading@10: lpc_p[n] = lpc[n] * pow_0_775[n]; yading@10: } yading@10: yading@10: ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s, yading@10: q->formant_mem + 10, 160, 10); yading@10: memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10); yading@10: ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10); yading@10: memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10); yading@10: yading@10: ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160); yading@10: yading@10: ff_adaptive_gain_control(samples, pole_out + 10, yading@10: avpriv_scalarproduct_float_c(q->formant_mem + 10, yading@10: q->formant_mem + 10, yading@10: 160), yading@10: 160, 0.9375, &q->postfilter_agc_mem); yading@10: } yading@10: yading@10: static int qcelp_decode_frame(AVCodecContext *avctx, void *data, yading@10: int *got_frame_ptr, AVPacket *avpkt) yading@10: { yading@10: const uint8_t *buf = avpkt->data; yading@10: int buf_size = avpkt->size; yading@10: QCELPContext *q = avctx->priv_data; yading@10: AVFrame *frame = data; yading@10: float *outbuffer; yading@10: int i, ret; yading@10: float quantized_lspf[10], lpc[10]; yading@10: float gain[16]; yading@10: float *formant_mem; yading@10: yading@10: /* get output buffer */ yading@10: frame->nb_samples = 160; yading@10: if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) yading@10: return ret; yading@10: outbuffer = (float *)frame->data[0]; yading@10: yading@10: if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) { yading@10: warn_insufficient_frame_quality(avctx, "bitrate cannot be determined."); yading@10: goto erasure; yading@10: } yading@10: yading@10: if (q->bitrate == RATE_OCTAVE && yading@10: (q->first16bits = AV_RB16(buf)) == 0xFFFF) { yading@10: warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on."); yading@10: goto erasure; yading@10: } yading@10: yading@10: if (q->bitrate > SILENCE) { yading@10: const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate]; yading@10: const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] + yading@10: qcelp_unpacking_bitmaps_lengths[q->bitrate]; yading@10: uint8_t *unpacked_data = (uint8_t *)&q->frame; yading@10: yading@10: init_get_bits(&q->gb, buf, 8 * buf_size); yading@10: yading@10: memset(&q->frame, 0, sizeof(QCELPFrame)); yading@10: yading@10: for (; bitmaps < bitmaps_end; bitmaps++) yading@10: unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos; yading@10: yading@10: // Check for erasures/blanks on rates 1, 1/4 and 1/8. yading@10: if (q->frame.reserved) { yading@10: warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area."); yading@10: goto erasure; yading@10: } yading@10: if (q->bitrate == RATE_QUARTER && yading@10: codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) { yading@10: warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed."); yading@10: goto erasure; yading@10: } yading@10: yading@10: if (q->bitrate >= RATE_HALF) { yading@10: for (i = 0; i < 4; i++) { yading@10: if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) { yading@10: warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter."); yading@10: goto erasure; yading@10: } yading@10: } yading@10: } yading@10: } yading@10: yading@10: decode_gain_and_index(q, gain); yading@10: compute_svector(q, gain, outbuffer); yading@10: yading@10: if (decode_lspf(q, quantized_lspf) < 0) { yading@10: warn_insufficient_frame_quality(avctx, "Badly received packets in frame."); yading@10: goto erasure; yading@10: } yading@10: yading@10: apply_pitch_filters(q, outbuffer); yading@10: yading@10: if (q->bitrate == I_F_Q) { yading@10: erasure: yading@10: q->bitrate = I_F_Q; yading@10: q->erasure_count++; yading@10: decode_gain_and_index(q, gain); yading@10: compute_svector(q, gain, outbuffer); yading@10: decode_lspf(q, quantized_lspf); yading@10: apply_pitch_filters(q, outbuffer); yading@10: } else yading@10: q->erasure_count = 0; yading@10: yading@10: formant_mem = q->formant_mem + 10; yading@10: for (i = 0; i < 4; i++) { yading@10: interpolate_lpc(q, quantized_lspf, lpc, i); yading@10: ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40, 10); yading@10: formant_mem += 40; yading@10: } yading@10: yading@10: // postfilter, as per TIA/EIA/IS-733 2.4.8.6 yading@10: postfilter(q, outbuffer, lpc); yading@10: yading@10: memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float)); yading@10: yading@10: memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf)); yading@10: q->prev_bitrate = q->bitrate; yading@10: yading@10: *got_frame_ptr = 1; yading@10: yading@10: return buf_size; yading@10: } yading@10: yading@10: AVCodec ff_qcelp_decoder = { yading@10: .name = "qcelp", yading@10: .type = AVMEDIA_TYPE_AUDIO, yading@10: .id = AV_CODEC_ID_QCELP, yading@10: .init = qcelp_decode_init, yading@10: .decode = qcelp_decode_frame, yading@10: .capabilities = CODEC_CAP_DR1, yading@10: .priv_data_size = sizeof(QCELPContext), yading@10: .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"), yading@10: };