yading@10: /* yading@10: * AMR narrowband decoder yading@10: * Copyright (c) 2006-2007 Robert Swain yading@10: * Copyright (c) 2009 Colin McQuillan 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: /** yading@10: * @file yading@10: * AMR narrowband decoder yading@10: * yading@10: * This decoder uses floats for simplicity and so is not bit-exact. One yading@10: * difference is that differences in phase can accumulate. The test sequences yading@10: * in 3GPP TS 26.074 can still be useful. yading@10: * yading@10: * - Comparing this file's output to the output of the ref decoder gives a yading@10: * PSNR of 30 to 80. Plotting the output samples shows a difference in yading@10: * phase in some areas. yading@10: * yading@10: * - Comparing both decoders against their input, this decoder gives a similar yading@10: * PSNR. If the test sequence homing frames are removed (this decoder does yading@10: * not detect them), the PSNR is at least as good as the reference on 140 yading@10: * out of 169 tests. yading@10: */ yading@10: yading@10: yading@10: #include 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 "libavutil/common.h" yading@10: #include "libavutil/avassert.h" yading@10: #include "celp_math.h" yading@10: #include "celp_filters.h" yading@10: #include "acelp_filters.h" yading@10: #include "acelp_vectors.h" yading@10: #include "acelp_pitch_delay.h" yading@10: #include "lsp.h" yading@10: #include "amr.h" yading@10: #include "internal.h" yading@10: yading@10: #include "amrnbdata.h" yading@10: yading@10: #define AMR_BLOCK_SIZE 160 ///< samples per frame yading@10: #define AMR_SAMPLE_BOUND 32768.0 ///< threshold for synthesis overflow yading@10: yading@10: /** yading@10: * Scale from constructed speech to [-1,1] yading@10: * yading@10: * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but yading@10: * upscales by two (section 6.2.2). yading@10: * yading@10: * Fundamentally, this scale is determined by energy_mean through yading@10: * the fixed vector contribution to the excitation vector. yading@10: */ yading@10: #define AMR_SAMPLE_SCALE (2.0 / 32768.0) yading@10: yading@10: /** Prediction factor for 12.2kbit/s mode */ yading@10: #define PRED_FAC_MODE_12k2 0.65 yading@10: yading@10: #define LSF_R_FAC (8000.0 / 32768.0) ///< LSF residual tables to Hertz yading@10: #define MIN_LSF_SPACING (50.0488 / 8000.0) ///< Ensures stability of LPC filter yading@10: #define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode yading@10: yading@10: /** Initial energy in dB. Also used for bad frames (unimplemented). */ yading@10: #define MIN_ENERGY -14.0 yading@10: yading@10: /** Maximum sharpening factor yading@10: * yading@10: * The specification says 0.8, which should be 13107, but the reference C code yading@10: * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in g729dec.c.) yading@10: */ yading@10: #define SHARP_MAX 0.79449462890625 yading@10: yading@10: /** Number of impulse response coefficients used for tilt factor */ yading@10: #define AMR_TILT_RESPONSE 22 yading@10: /** Tilt factor = 1st reflection coefficient * gamma_t */ yading@10: #define AMR_TILT_GAMMA_T 0.8 yading@10: /** Adaptive gain control factor used in post-filter */ yading@10: #define AMR_AGC_ALPHA 0.9 yading@10: yading@10: typedef struct AMRContext { yading@10: AMRNBFrame frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc) yading@10: uint8_t bad_frame_indicator; ///< bad frame ? 1 : 0 yading@10: enum Mode cur_frame_mode; yading@10: yading@10: int16_t prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe yading@10: double lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame yading@10: double prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame yading@10: yading@10: float lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing yading@10: float lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector yading@10: yading@10: float lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes yading@10: yading@10: uint8_t pitch_lag_int; ///< integer part of pitch lag from current subframe yading@10: yading@10: float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history yading@10: float *excitation; ///< pointer to the current excitation vector in excitation_buf yading@10: yading@10: float pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector yading@10: float fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames) yading@10: yading@10: float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes yading@10: float pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes yading@10: float fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes yading@10: yading@10: float beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX] yading@10: uint8_t diff_count; ///< the number of subframes for which diff has been above 0.65 yading@10: uint8_t hang_count; ///< the number of subframes since a hangover period started yading@10: yading@10: float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset" yading@10: uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none yading@10: uint8_t ir_filter_onset; ///< flag for impulse response filter strength yading@10: yading@10: float postfilter_mem[10]; ///< previous intermediate values in the formant filter yading@10: float tilt_mem; ///< previous input to tilt compensation filter yading@10: float postfilter_agc; ///< previous factor used for adaptive gain control yading@10: float high_pass_mem[2]; ///< previous intermediate values in the high-pass filter yading@10: yading@10: float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples yading@10: yading@10: ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs yading@10: ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs yading@10: CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs yading@10: CELPMContext celpm_ctx; ///< context for fixed point math operations yading@10: yading@10: } AMRContext; yading@10: yading@10: /** Double version of ff_weighted_vector_sumf() */ yading@10: static void weighted_vector_sumd(double *out, const double *in_a, yading@10: const double *in_b, double weight_coeff_a, yading@10: double weight_coeff_b, int length) yading@10: { yading@10: int i; yading@10: yading@10: for (i = 0; i < length; i++) yading@10: out[i] = weight_coeff_a * in_a[i] yading@10: + weight_coeff_b * in_b[i]; yading@10: } yading@10: yading@10: static av_cold int amrnb_decode_init(AVCodecContext *avctx) yading@10: { yading@10: AMRContext *p = avctx->priv_data; yading@10: int i; yading@10: yading@10: if (avctx->channels > 1) { yading@10: avpriv_report_missing_feature(avctx, "multi-channel AMR"); yading@10: return AVERROR_PATCHWELCOME; yading@10: } yading@10: yading@10: avctx->channels = 1; yading@10: avctx->channel_layout = AV_CH_LAYOUT_MONO; yading@10: if (!avctx->sample_rate) yading@10: avctx->sample_rate = 8000; yading@10: avctx->sample_fmt = AV_SAMPLE_FMT_FLT; yading@10: yading@10: // p->excitation always points to the same position in p->excitation_buf yading@10: p->excitation = &p->excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1]; yading@10: yading@10: for (i = 0; i < LP_FILTER_ORDER; i++) { yading@10: p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15); yading@10: p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15); yading@10: } yading@10: yading@10: for (i = 0; i < 4; i++) yading@10: p->prediction_error[i] = MIN_ENERGY; yading@10: yading@10: ff_acelp_filter_init(&p->acelpf_ctx); yading@10: ff_acelp_vectors_init(&p->acelpv_ctx); yading@10: ff_celp_filter_init(&p->celpf_ctx); yading@10: ff_celp_math_init(&p->celpm_ctx); yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: yading@10: /** yading@10: * Unpack an RFC4867 speech frame into the AMR frame mode and parameters. yading@10: * yading@10: * The order of speech bits is specified by 3GPP TS 26.101. yading@10: * yading@10: * @param p the context yading@10: * @param buf pointer to the input buffer yading@10: * @param buf_size size of the input buffer yading@10: * yading@10: * @return the frame mode yading@10: */ yading@10: static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf, yading@10: int buf_size) yading@10: { yading@10: enum Mode mode; yading@10: yading@10: // Decode the first octet. yading@10: mode = buf[0] >> 3 & 0x0F; // frame type yading@10: p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit yading@10: yading@10: if (mode >= N_MODES || buf_size < frame_sizes_nb[mode] + 1) { yading@10: return NO_DATA; yading@10: } yading@10: yading@10: if (mode < MODE_DTX) yading@10: ff_amr_bit_reorder((uint16_t *) &p->frame, sizeof(AMRNBFrame), buf + 1, yading@10: amr_unpacking_bitmaps_per_mode[mode]); yading@10: yading@10: return mode; yading@10: } yading@10: yading@10: yading@10: /// @name AMR pitch LPC coefficient decoding functions yading@10: /// @{ yading@10: yading@10: /** yading@10: * Interpolate the LSF vector (used for fixed gain smoothing). yading@10: * The interpolation is done over all four subframes even in MODE_12k2. yading@10: * yading@10: * @param[in] ctx The Context yading@10: * @param[in,out] lsf_q LSFs in [0,1] for each subframe yading@10: * @param[in] lsf_new New LSFs in [0,1] for subframe 4 yading@10: */ yading@10: static void interpolate_lsf(ACELPVContext *ctx, float lsf_q[4][LP_FILTER_ORDER], float *lsf_new) yading@10: { yading@10: int i; yading@10: yading@10: for (i = 0; i < 4; i++) yading@10: ctx->weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new, yading@10: 0.25 * (3 - i), 0.25 * (i + 1), yading@10: LP_FILTER_ORDER); yading@10: } yading@10: yading@10: /** yading@10: * Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector. yading@10: * yading@10: * @param p the context yading@10: * @param lsp output LSP vector yading@10: * @param lsf_no_r LSF vector without the residual vector added yading@10: * @param lsf_quantizer pointers to LSF dictionary tables yading@10: * @param quantizer_offset offset in tables yading@10: * @param sign for the 3 dictionary table yading@10: * @param update store data for computing the next frame's LSFs yading@10: */ yading@10: static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER], yading@10: const float lsf_no_r[LP_FILTER_ORDER], yading@10: const int16_t *lsf_quantizer[5], yading@10: const int quantizer_offset, yading@10: const int sign, const int update) yading@10: { yading@10: int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector yading@10: float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector yading@10: int i; yading@10: yading@10: for (i = 0; i < LP_FILTER_ORDER >> 1; i++) yading@10: memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset], yading@10: 2 * sizeof(*lsf_r)); yading@10: yading@10: if (sign) { yading@10: lsf_r[4] *= -1; yading@10: lsf_r[5] *= -1; yading@10: } yading@10: yading@10: if (update) yading@10: memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r)); yading@10: yading@10: for (i = 0; i < LP_FILTER_ORDER; i++) yading@10: lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0); yading@10: yading@10: ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER); yading@10: yading@10: if (update) yading@10: interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q); yading@10: yading@10: ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER); yading@10: } yading@10: yading@10: /** yading@10: * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors. yading@10: * yading@10: * @param p pointer to the AMRContext yading@10: */ yading@10: static void lsf2lsp_5(AMRContext *p) yading@10: { yading@10: const uint16_t *lsf_param = p->frame.lsf; yading@10: float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector yading@10: const int16_t *lsf_quantizer[5]; yading@10: int i; yading@10: yading@10: lsf_quantizer[0] = lsf_5_1[lsf_param[0]]; yading@10: lsf_quantizer[1] = lsf_5_2[lsf_param[1]]; yading@10: lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1]; yading@10: lsf_quantizer[3] = lsf_5_4[lsf_param[3]]; yading@10: lsf_quantizer[4] = lsf_5_5[lsf_param[4]]; yading@10: yading@10: for (i = 0; i < LP_FILTER_ORDER; i++) yading@10: lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i]; yading@10: yading@10: lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0); yading@10: lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1); yading@10: yading@10: // interpolate LSP vectors at subframes 1 and 3 yading@10: weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER); yading@10: weighted_vector_sumd(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER); yading@10: } yading@10: yading@10: /** yading@10: * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector. yading@10: * yading@10: * @param p pointer to the AMRContext yading@10: */ yading@10: static void lsf2lsp_3(AMRContext *p) yading@10: { yading@10: const uint16_t *lsf_param = p->frame.lsf; yading@10: int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector yading@10: float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector yading@10: const int16_t *lsf_quantizer; yading@10: int i, j; yading@10: yading@10: lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]]; yading@10: memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r)); yading@10: yading@10: lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)]; yading@10: memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r)); yading@10: yading@10: lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]]; yading@10: memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r)); yading@10: yading@10: // calculate mean-removed LSF vector and add mean yading@10: for (i = 0; i < LP_FILTER_ORDER; i++) yading@10: lsf_q[i] = (lsf_r[i] + p->prev_lsf_r[i] * pred_fac[i]) * (LSF_R_FAC / 8000.0) + lsf_3_mean[i] * (1.0 / 8000.0); yading@10: yading@10: ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER); yading@10: yading@10: // store data for computing the next frame's LSFs yading@10: interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q); yading@10: memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r)); yading@10: yading@10: ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER); yading@10: yading@10: // interpolate LSP vectors at subframes 1, 2 and 3 yading@10: for (i = 1; i <= 3; i++) yading@10: for(j = 0; j < LP_FILTER_ORDER; j++) yading@10: p->lsp[i-1][j] = p->prev_lsp_sub4[j] + yading@10: (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i; yading@10: } yading@10: yading@10: /// @} yading@10: yading@10: yading@10: /// @name AMR pitch vector decoding functions yading@10: /// @{ yading@10: yading@10: /** yading@10: * Like ff_decode_pitch_lag(), but with 1/6 resolution yading@10: */ yading@10: static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index, yading@10: const int prev_lag_int, const int subframe) yading@10: { yading@10: if (subframe == 0 || subframe == 2) { yading@10: if (pitch_index < 463) { yading@10: *lag_int = (pitch_index + 107) * 10923 >> 16; yading@10: *lag_frac = pitch_index - *lag_int * 6 + 105; yading@10: } else { yading@10: *lag_int = pitch_index - 368; yading@10: *lag_frac = 0; yading@10: } yading@10: } else { yading@10: *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1; yading@10: *lag_frac = pitch_index - *lag_int * 6 - 3; yading@10: *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2, yading@10: PITCH_DELAY_MAX - 9); yading@10: } yading@10: } yading@10: yading@10: static void decode_pitch_vector(AMRContext *p, yading@10: const AMRNBSubframe *amr_subframe, yading@10: const int subframe) yading@10: { yading@10: int pitch_lag_int, pitch_lag_frac; yading@10: enum Mode mode = p->cur_frame_mode; yading@10: yading@10: if (p->cur_frame_mode == MODE_12k2) { yading@10: decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac, yading@10: amr_subframe->p_lag, p->pitch_lag_int, yading@10: subframe); yading@10: } else yading@10: ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac, yading@10: amr_subframe->p_lag, yading@10: p->pitch_lag_int, subframe, yading@10: mode != MODE_4k75 && mode != MODE_5k15, yading@10: mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6)); yading@10: yading@10: p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t yading@10: yading@10: pitch_lag_frac <<= (p->cur_frame_mode != MODE_12k2); yading@10: yading@10: pitch_lag_int += pitch_lag_frac > 0; yading@10: yading@10: /* Calculate the pitch vector by interpolating the past excitation at the yading@10: pitch lag using a b60 hamming windowed sinc function. */ yading@10: p->acelpf_ctx.acelp_interpolatef(p->excitation, yading@10: p->excitation + 1 - pitch_lag_int, yading@10: ff_b60_sinc, 6, yading@10: pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0), yading@10: 10, AMR_SUBFRAME_SIZE); yading@10: yading@10: memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float)); yading@10: } yading@10: yading@10: /// @} yading@10: yading@10: yading@10: /// @name AMR algebraic code book (fixed) vector decoding functions yading@10: /// @{ yading@10: yading@10: /** yading@10: * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame. yading@10: */ yading@10: static void decode_10bit_pulse(int code, int pulse_position[8], yading@10: int i1, int i2, int i3) yading@10: { yading@10: // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of yading@10: // the 3 pulses and the upper 7 bits being coded in base 5 yading@10: const uint8_t *positions = base_five_table[code >> 3]; yading@10: pulse_position[i1] = (positions[2] << 1) + ( code & 1); yading@10: pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1); yading@10: pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1); yading@10: } yading@10: yading@10: /** yading@10: * Decode the algebraic codebook index to pulse positions and signs and yading@10: * construct the algebraic codebook vector for MODE_10k2. yading@10: * yading@10: * @param fixed_index positions of the eight pulses yading@10: * @param fixed_sparse pointer to the algebraic codebook vector yading@10: */ yading@10: static void decode_8_pulses_31bits(const int16_t *fixed_index, yading@10: AMRFixed *fixed_sparse) yading@10: { yading@10: int pulse_position[8]; yading@10: int i, temp; yading@10: yading@10: decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1); yading@10: decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5); yading@10: yading@10: // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of yading@10: // the 2 pulses and the upper 5 bits being coded in base 5 yading@10: temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5; yading@10: pulse_position[3] = temp % 5; yading@10: pulse_position[7] = temp / 5; yading@10: if (pulse_position[7] & 1) yading@10: pulse_position[3] = 4 - pulse_position[3]; yading@10: pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1); yading@10: pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1); yading@10: yading@10: fixed_sparse->n = 8; yading@10: for (i = 0; i < 4; i++) { yading@10: const int pos1 = (pulse_position[i] << 2) + i; yading@10: const int pos2 = (pulse_position[i + 4] << 2) + i; yading@10: const float sign = fixed_index[i] ? -1.0 : 1.0; yading@10: fixed_sparse->x[i ] = pos1; yading@10: fixed_sparse->x[i + 4] = pos2; yading@10: fixed_sparse->y[i ] = sign; yading@10: fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Decode the algebraic codebook index to pulse positions and signs, yading@10: * then construct the algebraic codebook vector. yading@10: * yading@10: * nb of pulses | bits encoding pulses yading@10: * For MODE_4k75 or MODE_5k15, 2 | 1-3, 4-6, 7 yading@10: * MODE_5k9, 2 | 1, 2-4, 5-6, 7-9 yading@10: * MODE_6k7, 3 | 1-3, 4, 5-7, 8, 9-11 yading@10: * MODE_7k4 or MODE_7k95, 4 | 1-3, 4-6, 7-9, 10, 11-13 yading@10: * yading@10: * @param fixed_sparse pointer to the algebraic codebook vector yading@10: * @param pulses algebraic codebook indexes yading@10: * @param mode mode of the current frame yading@10: * @param subframe current subframe number yading@10: */ yading@10: static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses, yading@10: const enum Mode mode, const int subframe) yading@10: { yading@10: av_assert1(MODE_4k75 <= (signed)mode && mode <= MODE_12k2); yading@10: yading@10: if (mode == MODE_12k2) { yading@10: ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3); yading@10: } else if (mode == MODE_10k2) { yading@10: decode_8_pulses_31bits(pulses, fixed_sparse); yading@10: } else { yading@10: int *pulse_position = fixed_sparse->x; yading@10: int i, pulse_subset; yading@10: const int fixed_index = pulses[0]; yading@10: yading@10: if (mode <= MODE_5k15) { yading@10: pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1); yading@10: pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset]; yading@10: pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1]; yading@10: fixed_sparse->n = 2; yading@10: } else if (mode == MODE_5k9) { yading@10: pulse_subset = ((fixed_index & 1) << 1) + 1; yading@10: pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset; yading@10: pulse_subset = (fixed_index >> 4) & 3; yading@10: pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0); yading@10: fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2; yading@10: } else if (mode == MODE_6k7) { yading@10: pulse_position[0] = (fixed_index & 7) * 5; yading@10: pulse_subset = (fixed_index >> 2) & 2; yading@10: pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1; yading@10: pulse_subset = (fixed_index >> 6) & 2; yading@10: pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2; yading@10: fixed_sparse->n = 3; yading@10: } else { // mode <= MODE_7k95 yading@10: pulse_position[0] = gray_decode[ fixed_index & 7]; yading@10: pulse_position[1] = gray_decode[(fixed_index >> 3) & 7] + 1; yading@10: pulse_position[2] = gray_decode[(fixed_index >> 6) & 7] + 2; yading@10: pulse_subset = (fixed_index >> 9) & 1; yading@10: pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3; yading@10: fixed_sparse->n = 4; yading@10: } yading@10: for (i = 0; i < fixed_sparse->n; i++) yading@10: fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2) yading@10: * yading@10: * @param p the context yading@10: * @param subframe unpacked amr subframe yading@10: * @param mode mode of the current frame yading@10: * @param fixed_sparse sparse respresentation of the fixed vector yading@10: */ yading@10: static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode, yading@10: AMRFixed *fixed_sparse) yading@10: { yading@10: // The spec suggests the current pitch gain is always used, but in other yading@10: // modes the pitch and codebook gains are joinly quantized (sec 5.8.2) yading@10: // so the codebook gain cannot depend on the quantized pitch gain. yading@10: if (mode == MODE_12k2) yading@10: p->beta = FFMIN(p->pitch_gain[4], 1.0); yading@10: yading@10: fixed_sparse->pitch_lag = p->pitch_lag_int; yading@10: fixed_sparse->pitch_fac = p->beta; yading@10: yading@10: // Save pitch sharpening factor for the next subframe yading@10: // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from yading@10: // the fact that the gains for two subframes are jointly quantized. yading@10: if (mode != MODE_4k75 || subframe & 1) yading@10: p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX); yading@10: } yading@10: /// @} yading@10: yading@10: yading@10: /// @name AMR gain decoding functions yading@10: /// @{ yading@10: yading@10: /** yading@10: * fixed gain smoothing yading@10: * Note that where the spec specifies the "spectrum in the q domain" yading@10: * in section 6.1.4, in fact frequencies should be used. yading@10: * yading@10: * @param p the context yading@10: * @param lsf LSFs for the current subframe, in the range [0,1] yading@10: * @param lsf_avg averaged LSFs yading@10: * @param mode mode of the current frame yading@10: * yading@10: * @return fixed gain smoothed yading@10: */ yading@10: static float fixed_gain_smooth(AMRContext *p , const float *lsf, yading@10: const float *lsf_avg, const enum Mode mode) yading@10: { yading@10: float diff = 0.0; yading@10: int i; yading@10: yading@10: for (i = 0; i < LP_FILTER_ORDER; i++) yading@10: diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i]; yading@10: yading@10: // If diff is large for ten subframes, disable smoothing for a 40-subframe yading@10: // hangover period. yading@10: p->diff_count++; yading@10: if (diff <= 0.65) yading@10: p->diff_count = 0; yading@10: yading@10: if (p->diff_count > 10) { yading@10: p->hang_count = 0; yading@10: p->diff_count--; // don't let diff_count overflow yading@10: } yading@10: yading@10: if (p->hang_count < 40) { yading@10: p->hang_count++; yading@10: } else if (mode < MODE_7k4 || mode == MODE_10k2) { yading@10: const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0); yading@10: const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] + yading@10: p->fixed_gain[2] + p->fixed_gain[3] + yading@10: p->fixed_gain[4]) * 0.2; yading@10: return smoothing_factor * p->fixed_gain[4] + yading@10: (1.0 - smoothing_factor) * fixed_gain_mean; yading@10: } yading@10: return p->fixed_gain[4]; yading@10: } yading@10: yading@10: /** yading@10: * Decode pitch gain and fixed gain factor (part of section 6.1.3). yading@10: * yading@10: * @param p the context yading@10: * @param amr_subframe unpacked amr subframe yading@10: * @param mode mode of the current frame yading@10: * @param subframe current subframe number yading@10: * @param fixed_gain_factor decoded gain correction factor yading@10: */ yading@10: static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe, yading@10: const enum Mode mode, const int subframe, yading@10: float *fixed_gain_factor) yading@10: { yading@10: if (mode == MODE_12k2 || mode == MODE_7k95) { yading@10: p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ] yading@10: * (1.0 / 16384.0); yading@10: *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain] yading@10: * (1.0 / 2048.0); yading@10: } else { yading@10: const uint16_t *gains; yading@10: yading@10: if (mode >= MODE_6k7) { yading@10: gains = gains_high[amr_subframe->p_gain]; yading@10: } else if (mode >= MODE_5k15) { yading@10: gains = gains_low [amr_subframe->p_gain]; yading@10: } else { yading@10: // gain index is only coded in subframes 0,2 for MODE_4k75 yading@10: gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)]; yading@10: } yading@10: yading@10: p->pitch_gain[4] = gains[0] * (1.0 / 16384.0); yading@10: *fixed_gain_factor = gains[1] * (1.0 / 4096.0); yading@10: } yading@10: } yading@10: yading@10: /// @} yading@10: yading@10: yading@10: /// @name AMR preprocessing functions yading@10: /// @{ yading@10: yading@10: /** yading@10: * Circularly convolve a sparse fixed vector with a phase dispersion impulse yading@10: * response filter (D.6.2 of G.729 and 6.1.5 of AMR). yading@10: * yading@10: * @param out vector with filter applied yading@10: * @param in source vector yading@10: * @param filter phase filter coefficients yading@10: * yading@10: * out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] } yading@10: */ yading@10: static void apply_ir_filter(float *out, const AMRFixed *in, yading@10: const float *filter) yading@10: { yading@10: float filter1[AMR_SUBFRAME_SIZE], ///< filters at pitch lag*1 and *2 yading@10: filter2[AMR_SUBFRAME_SIZE]; yading@10: int lag = in->pitch_lag; yading@10: float fac = in->pitch_fac; yading@10: int i; yading@10: yading@10: if (lag < AMR_SUBFRAME_SIZE) { yading@10: ff_celp_circ_addf(filter1, filter, filter, lag, fac, yading@10: AMR_SUBFRAME_SIZE); yading@10: yading@10: if (lag < AMR_SUBFRAME_SIZE >> 1) yading@10: ff_celp_circ_addf(filter2, filter, filter1, lag, fac, yading@10: AMR_SUBFRAME_SIZE); yading@10: } yading@10: yading@10: memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE); yading@10: for (i = 0; i < in->n; i++) { yading@10: int x = in->x[i]; yading@10: float y = in->y[i]; yading@10: const float *filterp; yading@10: yading@10: if (x >= AMR_SUBFRAME_SIZE - lag) { yading@10: filterp = filter; yading@10: } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) { yading@10: filterp = filter1; yading@10: } else yading@10: filterp = filter2; yading@10: yading@10: ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE); yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Reduce fixed vector sparseness by smoothing with one of three IR filters. yading@10: * Also know as "adaptive phase dispersion". yading@10: * yading@10: * This implements 3GPP TS 26.090 section 6.1(5). yading@10: * yading@10: * @param p the context yading@10: * @param fixed_sparse algebraic codebook vector yading@10: * @param fixed_vector unfiltered fixed vector yading@10: * @param fixed_gain smoothed gain yading@10: * @param out space for modified vector if necessary yading@10: */ yading@10: static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse, yading@10: const float *fixed_vector, yading@10: float fixed_gain, float *out) yading@10: { yading@10: int ir_filter_nr; yading@10: yading@10: if (p->pitch_gain[4] < 0.6) { yading@10: ir_filter_nr = 0; // strong filtering yading@10: } else if (p->pitch_gain[4] < 0.9) { yading@10: ir_filter_nr = 1; // medium filtering yading@10: } else yading@10: ir_filter_nr = 2; // no filtering yading@10: yading@10: // detect 'onset' yading@10: if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) { yading@10: p->ir_filter_onset = 2; yading@10: } else if (p->ir_filter_onset) yading@10: p->ir_filter_onset--; yading@10: yading@10: if (!p->ir_filter_onset) { yading@10: int i, count = 0; yading@10: yading@10: for (i = 0; i < 5; i++) yading@10: if (p->pitch_gain[i] < 0.6) yading@10: count++; yading@10: if (count > 2) yading@10: ir_filter_nr = 0; yading@10: yading@10: if (ir_filter_nr > p->prev_ir_filter_nr + 1) yading@10: ir_filter_nr--; yading@10: } else if (ir_filter_nr < 2) yading@10: ir_filter_nr++; yading@10: yading@10: // Disable filtering for very low level of fixed_gain. yading@10: // Note this step is not specified in the technical description but is in yading@10: // the reference source in the function Ph_disp. yading@10: if (fixed_gain < 5.0) yading@10: ir_filter_nr = 2; yading@10: yading@10: if (p->cur_frame_mode != MODE_7k4 && p->cur_frame_mode < MODE_10k2 yading@10: && ir_filter_nr < 2) { yading@10: apply_ir_filter(out, fixed_sparse, yading@10: (p->cur_frame_mode == MODE_7k95 ? yading@10: ir_filters_lookup_MODE_7k95 : yading@10: ir_filters_lookup)[ir_filter_nr]); yading@10: fixed_vector = out; yading@10: } yading@10: yading@10: // update ir filter strength history yading@10: p->prev_ir_filter_nr = ir_filter_nr; yading@10: p->prev_sparse_fixed_gain = fixed_gain; yading@10: yading@10: return fixed_vector; yading@10: } yading@10: yading@10: /// @} yading@10: yading@10: yading@10: /// @name AMR synthesis functions yading@10: /// @{ yading@10: yading@10: /** yading@10: * Conduct 10th order linear predictive coding synthesis. yading@10: * yading@10: * @param p pointer to the AMRContext yading@10: * @param lpc pointer to the LPC coefficients yading@10: * @param fixed_gain fixed codebook gain for synthesis yading@10: * @param fixed_vector algebraic codebook vector yading@10: * @param samples pointer to the output speech samples yading@10: * @param overflow 16-bit overflow flag yading@10: */ yading@10: static int synthesis(AMRContext *p, float *lpc, yading@10: float fixed_gain, const float *fixed_vector, yading@10: float *samples, uint8_t overflow) yading@10: { yading@10: int i; yading@10: float excitation[AMR_SUBFRAME_SIZE]; yading@10: yading@10: // if an overflow has been detected, the pitch vector is scaled down by a yading@10: // factor of 4 yading@10: if (overflow) yading@10: for (i = 0; i < AMR_SUBFRAME_SIZE; i++) yading@10: p->pitch_vector[i] *= 0.25; yading@10: yading@10: p->acelpv_ctx.weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector, yading@10: p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE); yading@10: yading@10: // emphasize pitch vector contribution yading@10: if (p->pitch_gain[4] > 0.5 && !overflow) { yading@10: float energy = p->celpm_ctx.dot_productf(excitation, excitation, yading@10: AMR_SUBFRAME_SIZE); yading@10: float pitch_factor = yading@10: p->pitch_gain[4] * yading@10: (p->cur_frame_mode == MODE_12k2 ? yading@10: 0.25 * FFMIN(p->pitch_gain[4], 1.0) : yading@10: 0.5 * FFMIN(p->pitch_gain[4], SHARP_MAX)); yading@10: yading@10: for (i = 0; i < AMR_SUBFRAME_SIZE; i++) yading@10: excitation[i] += pitch_factor * p->pitch_vector[i]; yading@10: yading@10: ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy, yading@10: AMR_SUBFRAME_SIZE); yading@10: } yading@10: yading@10: p->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation, yading@10: AMR_SUBFRAME_SIZE, yading@10: LP_FILTER_ORDER); yading@10: yading@10: // detect overflow yading@10: for (i = 0; i < AMR_SUBFRAME_SIZE; i++) yading@10: if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) { yading@10: return 1; yading@10: } yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: /// @} yading@10: yading@10: yading@10: /// @name AMR update functions yading@10: /// @{ yading@10: yading@10: /** yading@10: * Update buffers and history at the end of decoding a subframe. yading@10: * yading@10: * @param p pointer to the AMRContext yading@10: */ yading@10: static void update_state(AMRContext *p) yading@10: { yading@10: memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0])); yading@10: yading@10: memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE], yading@10: (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float)); yading@10: yading@10: memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float)); yading@10: memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float)); yading@10: yading@10: memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE], yading@10: LP_FILTER_ORDER * sizeof(float)); yading@10: } yading@10: yading@10: /// @} yading@10: yading@10: yading@10: /// @name AMR Postprocessing functions yading@10: /// @{ yading@10: yading@10: /** yading@10: * Get the tilt factor of a formant filter from its transfer function yading@10: * yading@10: * @param p The Context yading@10: * @param lpc_n LP_FILTER_ORDER coefficients of the numerator yading@10: * @param lpc_d LP_FILTER_ORDER coefficients of the denominator yading@10: */ yading@10: static float tilt_factor(AMRContext *p, float *lpc_n, float *lpc_d) yading@10: { yading@10: float rh0, rh1; // autocorrelation at lag 0 and 1 yading@10: yading@10: // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf yading@10: float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 }; yading@10: float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response yading@10: yading@10: hf[0] = 1.0; yading@10: memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER); yading@10: p->celpf_ctx.celp_lp_synthesis_filterf(hf, lpc_d, hf, yading@10: AMR_TILT_RESPONSE, yading@10: LP_FILTER_ORDER); yading@10: yading@10: rh0 = p->celpm_ctx.dot_productf(hf, hf, AMR_TILT_RESPONSE); yading@10: rh1 = p->celpm_ctx.dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1); yading@10: yading@10: // The spec only specifies this check for 12.2 and 10.2 kbit/s yading@10: // modes. But in the ref source the tilt is always non-negative. yading@10: return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0; yading@10: } yading@10: yading@10: /** yading@10: * Perform adaptive post-filtering to enhance the quality of the speech. yading@10: * See section 6.2.1. yading@10: * yading@10: * @param p pointer to the AMRContext yading@10: * @param lpc interpolated LP coefficients for this subframe yading@10: * @param buf_out output of the filter yading@10: */ yading@10: static void postfilter(AMRContext *p, float *lpc, float *buf_out) yading@10: { yading@10: int i; yading@10: float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input yading@10: yading@10: float speech_gain = p->celpm_ctx.dot_productf(samples, samples, yading@10: AMR_SUBFRAME_SIZE); yading@10: yading@10: float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter yading@10: const float *gamma_n, *gamma_d; // Formant filter factor table yading@10: float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients yading@10: yading@10: if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) { yading@10: gamma_n = ff_pow_0_7; yading@10: gamma_d = ff_pow_0_75; yading@10: } else { yading@10: gamma_n = ff_pow_0_55; yading@10: gamma_d = ff_pow_0_7; yading@10: } yading@10: yading@10: for (i = 0; i < LP_FILTER_ORDER; i++) { yading@10: lpc_n[i] = lpc[i] * gamma_n[i]; yading@10: lpc_d[i] = lpc[i] * gamma_d[i]; yading@10: } yading@10: yading@10: memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER); yading@10: p->celpf_ctx.celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples, yading@10: AMR_SUBFRAME_SIZE, LP_FILTER_ORDER); yading@10: memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE, yading@10: sizeof(float) * LP_FILTER_ORDER); yading@10: yading@10: p->celpf_ctx.celp_lp_zero_synthesis_filterf(buf_out, lpc_n, yading@10: pole_out + LP_FILTER_ORDER, yading@10: AMR_SUBFRAME_SIZE, LP_FILTER_ORDER); yading@10: yading@10: ff_tilt_compensation(&p->tilt_mem, tilt_factor(p, lpc_n, lpc_d), buf_out, yading@10: AMR_SUBFRAME_SIZE); yading@10: yading@10: ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE, yading@10: AMR_AGC_ALPHA, &p->postfilter_agc); yading@10: } yading@10: yading@10: /// @} yading@10: yading@10: static int amrnb_decode_frame(AVCodecContext *avctx, void *data, yading@10: int *got_frame_ptr, AVPacket *avpkt) yading@10: { yading@10: yading@10: AMRContext *p = avctx->priv_data; // pointer to private data yading@10: AVFrame *frame = data; yading@10: const uint8_t *buf = avpkt->data; yading@10: int buf_size = avpkt->size; yading@10: float *buf_out; // pointer to the output data buffer yading@10: int i, subframe, ret; yading@10: float fixed_gain_factor; yading@10: AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing yading@10: float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing yading@10: float synth_fixed_gain; // the fixed gain that synthesis should use yading@10: const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use yading@10: yading@10: /* get output buffer */ yading@10: frame->nb_samples = AMR_BLOCK_SIZE; yading@10: if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) yading@10: return ret; yading@10: buf_out = (float *)frame->data[0]; yading@10: yading@10: p->cur_frame_mode = unpack_bitstream(p, buf, buf_size); yading@10: if (p->cur_frame_mode == NO_DATA) { yading@10: av_log(avctx, AV_LOG_ERROR, "Corrupt bitstream\n"); yading@10: return AVERROR_INVALIDDATA; yading@10: } yading@10: if (p->cur_frame_mode == MODE_DTX) { yading@10: avpriv_report_missing_feature(avctx, "dtx mode"); yading@10: av_log(avctx, AV_LOG_INFO, "Note: libopencore_amrnb supports dtx\n"); yading@10: return AVERROR_PATCHWELCOME; yading@10: } yading@10: yading@10: if (p->cur_frame_mode == MODE_12k2) { yading@10: lsf2lsp_5(p); yading@10: } else yading@10: lsf2lsp_3(p); yading@10: yading@10: for (i = 0; i < 4; i++) yading@10: ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5); yading@10: yading@10: for (subframe = 0; subframe < 4; subframe++) { yading@10: const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe]; yading@10: yading@10: decode_pitch_vector(p, amr_subframe, subframe); yading@10: yading@10: decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses, yading@10: p->cur_frame_mode, subframe); yading@10: yading@10: // The fixed gain (section 6.1.3) depends on the fixed vector yading@10: // (section 6.1.2), but the fixed vector calculation uses yading@10: // pitch sharpening based on the on the pitch gain (section 6.1.3). yading@10: // So the correct order is: pitch gain, pitch sharpening, fixed gain. yading@10: decode_gains(p, amr_subframe, p->cur_frame_mode, subframe, yading@10: &fixed_gain_factor); yading@10: yading@10: pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse); yading@10: yading@10: if (fixed_sparse.pitch_lag == 0) { yading@10: av_log(avctx, AV_LOG_ERROR, "The file is corrupted, pitch_lag = 0 is not allowed\n"); yading@10: return AVERROR_INVALIDDATA; yading@10: } yading@10: ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0, yading@10: AMR_SUBFRAME_SIZE); yading@10: yading@10: p->fixed_gain[4] = yading@10: ff_amr_set_fixed_gain(fixed_gain_factor, yading@10: p->celpm_ctx.dot_productf(p->fixed_vector, yading@10: p->fixed_vector, yading@10: AMR_SUBFRAME_SIZE) / yading@10: AMR_SUBFRAME_SIZE, yading@10: p->prediction_error, yading@10: energy_mean[p->cur_frame_mode], energy_pred_fac); yading@10: yading@10: // The excitation feedback is calculated without any processing such yading@10: // as fixed gain smoothing. This isn't mentioned in the specification. yading@10: for (i = 0; i < AMR_SUBFRAME_SIZE; i++) yading@10: p->excitation[i] *= p->pitch_gain[4]; yading@10: ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4], yading@10: AMR_SUBFRAME_SIZE); yading@10: yading@10: // In the ref decoder, excitation is stored with no fractional bits. yading@10: // This step prevents buzz in silent periods. The ref encoder can yading@10: // emit long sequences with pitch factor greater than one. This yading@10: // creates unwanted feedback if the excitation vector is nonzero. yading@10: // (e.g. test sequence T19_795.COD in 3GPP TS 26.074) yading@10: for (i = 0; i < AMR_SUBFRAME_SIZE; i++) yading@10: p->excitation[i] = truncf(p->excitation[i]); yading@10: yading@10: // Smooth fixed gain. yading@10: // The specification is ambiguous, but in the reference source, the yading@10: // smoothed value is NOT fed back into later fixed gain smoothing. yading@10: synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe], yading@10: p->lsf_avg, p->cur_frame_mode); yading@10: yading@10: synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector, yading@10: synth_fixed_gain, spare_vector); yading@10: yading@10: if (synthesis(p, p->lpc[subframe], synth_fixed_gain, yading@10: synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0)) yading@10: // overflow detected -> rerun synthesis scaling pitch vector down yading@10: // by a factor of 4, skipping pitch vector contribution emphasis yading@10: // and adaptive gain control yading@10: synthesis(p, p->lpc[subframe], synth_fixed_gain, yading@10: synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1); yading@10: yading@10: postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE); yading@10: yading@10: // update buffers and history yading@10: ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE); yading@10: update_state(p); yading@10: } yading@10: yading@10: p->acelpf_ctx.acelp_apply_order_2_transfer_function(buf_out, yading@10: buf_out, highpass_zeros, yading@10: highpass_poles, yading@10: highpass_gain * AMR_SAMPLE_SCALE, yading@10: p->high_pass_mem, AMR_BLOCK_SIZE); yading@10: yading@10: /* Update averaged lsf vector (used for fixed gain smoothing). yading@10: * yading@10: * Note that lsf_avg should not incorporate the current frame's LSFs yading@10: * for fixed_gain_smooth. yading@10: * The specification has an incorrect formula: the reference decoder uses yading@10: * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */ yading@10: p->acelpv_ctx.weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3], yading@10: 0.84, 0.16, LP_FILTER_ORDER); yading@10: yading@10: *got_frame_ptr = 1; yading@10: yading@10: /* return the amount of bytes consumed if everything was OK */ yading@10: return frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC yading@10: } yading@10: yading@10: yading@10: AVCodec ff_amrnb_decoder = { yading@10: .name = "amrnb", yading@10: .type = AVMEDIA_TYPE_AUDIO, yading@10: .id = AV_CODEC_ID_AMR_NB, yading@10: .priv_data_size = sizeof(AMRContext), yading@10: .init = amrnb_decode_init, yading@10: .decode = amrnb_decode_frame, yading@10: .capabilities = CODEC_CAP_DR1, yading@10: .long_name = NULL_IF_CONFIG_SMALL("AMR-NB (Adaptive Multi-Rate NarrowBand)"), yading@10: .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT, yading@10: AV_SAMPLE_FMT_NONE }, yading@10: };