Commit 6458e11050 for ffmpeg
commit 6458e110505776ceeeab51953633f3a4c91dc37a
Author: Jun Zhao <barryjzhao@tencent.com>
Date: Sat Oct 10 16:13:16 2026 +0800
avcodec/amrnbdec: conceal lost speech frames
During speech, NO_DATA frames were emitted as digital silence,
while speech frames marked bad were decoded without concealment.
TS 26.093 clause 5.2.3 requires the substitution and muting
described in TS 26.091.
Follow TS 26.073 in floating point: a bad-frame state machine,
pseudo-random replacement of a NO_DATA frame, LSF concealment
towards the mean, median gain attenuation, and gradual pitch-lag
degradation. Update the NO_DATA FATE checks from silence to the
concealed output levels.
Speech outside a silent period is unchanged.
Signed-off-by: Jun Zhao <barryjzhao@tencent.com>
diff --git a/libavcodec/amrnbdec.c b/libavcodec/amrnbdec.c
index ecd138bf28..dc5a1dd842 100644
--- a/libavcodec/amrnbdec.c
+++ b/libavcodec/amrnbdec.c
@@ -81,6 +81,19 @@
#define MIN_LSF_SPACING (50.0488 / 8000.0) ///< Ensures stability of LPC filter
#define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode
+/**
+ * Weights of the past LSFs and of their mean when concealing a lost frame,
+ * ALPHA and ONE_ALPHA of 3GPP TS 26.073. The 12.2 kbit/s mode uses a
+ * different pair than all the other modes do:
+ *
+ * D_plsf_3() d_plsf_3.c:51-52 29491 / 3277 (0.90 / 0.10)
+ * D_plsf_5() d_plsf_5.c:49-50 31128 / 1639 (0.95 / 0.05)
+ */
+#define LSF_ALPHA (29491.0f / 32768.0f)
+#define LSF_ONE_ALPHA ( 3277.0f / 32768.0f)
+#define LSF_ALPHA_5 (31128.0f / 32768.0f)
+#define LSF_ONE_ALPHA_5 ( 1639.0f / 32768.0f)
+
/** Initial energy in dB. Also used for bad frames (unimplemented). */
#define MIN_ENERGY -14.0
@@ -192,8 +205,12 @@ typedef struct AMRContext {
int valid_data; ///< the CN parameters of this frame may be used
int data_updated; ///< CN parameters have been renewed at least once
int prev_bf; ///< the previous frame could not be used
- float prev_pitch_gain; ///< pitch gain of the last usable frame
- float prev_fixed_gain; ///< fixed gain of the last usable frame
+
+ /* Bad frame handling and error concealment state, 3GPP TS 26.091 */
+ int bfh_state; ///< bad frame handling state (0..6)
+ uint32_t no_data_seed; ///< pseudo-random source for lost frames
+ float ec_gain_pbuf[5], ec_gain_past_pit, ec_gain_prev_pit;
+ float ec_gain_gbuf[5], ec_gain_past_code, ec_gain_prev_code;
} AMRContext;
typedef struct AMRChannelsContext {
@@ -212,6 +229,30 @@ static void weighted_vector_sumd(double *out, const double *in_a,
+ weight_coeff_b * in_b[i];
}
+/**
+ * Reset the bad frame handling and gain concealment states, matching
+ * Decoder_amr_reset() and ec_gain_*_reset() in 3GPP TS 26.073.
+ */
+static void ec_reset(AMRContext *p)
+{
+ int i;
+
+ p->bfh_state = 0;
+ p->no_data_seed = 21845;
+ /* ec_gain_pitch_reset() stores 1640 in Q14. ec_gain_code_reset() stores 1
+ in Q1 (0.5). In this floating-point decoder p->excitation[] uses unit
+ pulse amplitude and is truncated with truncf(), so 1.0f is used as an
+ intentional unit-pulse floor rather than the 0.5 fixed-point scale. */
+ for (i = 0; i < 5; i++) {
+ p->ec_gain_pbuf[i] = 1640.0f / 16384.0f;
+ p->ec_gain_gbuf[i] = 1.0f;
+ }
+ p->ec_gain_past_pit = 0.0f;
+ p->ec_gain_prev_pit = 1.0f;
+ p->ec_gain_past_code = 0.0f;
+ p->ec_gain_prev_code = 1.0f;
+}
+
static av_cold int amrnb_decode_init(AVCodecContext *avctx)
{
AMRChannelsContext *s = avctx->priv_data;
@@ -252,8 +293,8 @@ static av_cold int amrnb_decode_init(AVCodecContext *avctx)
p->hangover_count = DTX_HANG_CONST;
p->dec_ana_elapsed = 32767;
p->prev_bf = 0;
- p->prev_pitch_gain = 0.0f;
- p->prev_fixed_gain = 0.0f;
+
+ ec_reset(p);
for (i = 0; i < LP_FILTER_ORDER; i++)
p->cn_lsp[i] = p->cn_lsp_old[i] = cn_lsp_init[i];
@@ -376,14 +417,44 @@ static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER],
* Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors.
*
* @param p pointer to the AMRContext
+ * @param bfi the frame is unusable, conceal the LSF vector
*/
-static void lsf2lsp_5(AMRContext *p)
+static void lsf2lsp_5(AMRContext *p, int bfi)
{
const uint16_t *lsf_param = p->frame.lsf;
float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector
const int16_t *lsf_quantizer[5];
+ float lsf_q[LP_FILTER_ORDER];
int i;
+ if (bfi) {
+ /* Use the past LSFs slightly shifted towards their mean, and update
+ the prediction residual, see D_plsf_5() in 3GPP TS 26.073
+ (d_plsf_5.c:88-115). The LSFs are kept in Hertz while the
+ residual prev_lsf_r is in the codebook domain (LSF_R_FAC). */
+ for (i = 0; i < LP_FILTER_ORDER; i++)
+ lsf_q[i] = (LSF_ALPHA_5 * p->lsf_q[3][i] * 8000.0f +
+ LSF_ONE_ALPHA_5 * lsf_5_mean[i]) * (1.0f / 8000.0f);
+
+ for (i = 0; i < LP_FILTER_ORDER; i++)
+ p->prev_lsf_r[i] = av_clip_int16(lrintf(
+ (lsf_q[i] * 8000.0f -
+ (lsf_5_mean[i] + PRED_FAC_MODE_12k2 * p->prev_lsf_r[i] *
+ LSF_R_FAC)) / LSF_R_FAC));
+
+ ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
+ interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
+ ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
+
+ /* the first and second LSP vectors are concealed identically */
+ memcpy(p->lsp[1], p->lsp[3], sizeof(p->lsp[1]));
+ weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5,
+ LP_FILTER_ORDER);
+ weighted_vector_sumd(p->lsp[2], p->lsp[1], p->lsp[3], 0.5, 0.5,
+ LP_FILTER_ORDER);
+ return;
+ }
+
lsf_quantizer[0] = lsf_5_1[lsf_param[0]];
lsf_quantizer[1] = lsf_5_2[lsf_param[1]];
lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1];
@@ -405,15 +476,44 @@ static void lsf2lsp_5(AMRContext *p)
* Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector.
*
* @param p pointer to the AMRContext
+ * @param bfi the frame is unusable, conceal the LSF vector
*/
-static void lsf2lsp_3(AMRContext *p)
+static void lsf2lsp_3(AMRContext *p, int bfi)
{
const uint16_t *lsf_param = p->frame.lsf;
- int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
- const int16_t *lsf_quantizer;
int i, j;
+ if (bfi) {
+ /* Use the past LSFs slightly shifted towards their mean, and update
+ the prediction residual, see D_plsf_3() in 3GPP TS 26.073
+ (d_plsf_3.c:84-115). The LSFs are kept in Hertz while the residual
+ prev_lsf_r is in the codebook domain (LSF_R_FAC). */
+ for (i = 0; i < LP_FILTER_ORDER; i++)
+ lsf_q[i] = (LSF_ALPHA * p->lsf_q[3][i] * 8000.0f +
+ LSF_ONE_ALPHA * lsf_3_mean[i]) * (1.0f / 8000.0f);
+
+ for (i = 0; i < LP_FILTER_ORDER; i++)
+ p->prev_lsf_r[i] = av_clip_int16(lrintf(
+ (lsf_q[i] * 8000.0f -
+ (lsf_3_mean[i] + pred_fac[i] * p->prev_lsf_r[i] *
+ LSF_R_FAC)) / LSF_R_FAC));
+
+ ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
+ interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
+ ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
+
+ // interpolate LSP vectors at subframes 1, 2 and 3
+ for (i = 1; i <= 3; i++)
+ for (j = 0; j < LP_FILTER_ORDER; j++)
+ p->lsp[i-1][j] = p->prev_lsp_sub4[j] +
+ (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i;
+ return;
+ }
+
+ int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
+ const int16_t *lsf_quantizer;
+
lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]];
memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r));
@@ -472,12 +572,22 @@ static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index,
static void decode_pitch_vector(AMRContext *p,
const AMRNBSubframe *amr_subframe,
- const int subframe)
+ const int subframe, const int bfi)
{
int pitch_lag_int, pitch_lag_frac;
enum Mode mode = p->cur_frame_mode;
- if (p->cur_frame_mode == MODE_12k2) {
+ if (bfi) {
+ /* Graceful pitch degradation for a lost frame, see Decoder_amr() in
+ 3GPP TS 26.073 (dec_amr.c:527-548 for the delta-coded modes and
+ dec_amr.c:559-568 for 12.2kbit/s). The integer lag keeps growing
+ until it reaches its maximum instead of being decoded. */
+ if (mode != MODE_12k2)
+ pitch_lag_int = FFMIN(p->pitch_lag_int + 1, PITCH_DELAY_MAX);
+ else
+ pitch_lag_int = p->pitch_lag_int;
+ pitch_lag_frac = 0;
+ } else if (p->cur_frame_mode == MODE_12k2) {
decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac,
amr_subframe->p_lag, p->pitch_lag_int,
subframe);
@@ -738,6 +848,149 @@ static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe,
/// @}
+/// @name AMR error concealment functions
+/// @{
+
+/**
+ * Median of the last five gain values, equivalent to gmed_n() in
+ * 3GPP TS 26.073 (gmed_n.c:57-84).
+ */
+static float ec_gain_median(const float *buf)
+{
+ float tmp[5];
+ int i, j;
+
+ memcpy(tmp, buf, sizeof(tmp));
+
+ for (i = 1; i < 5; i++) {
+ float v = tmp[i];
+
+ for (j = i - 1; j >= 0 && tmp[j] > v; j--)
+ tmp[j + 1] = tmp[j];
+ tmp[j + 1] = v;
+ }
+
+ return tmp[2];
+}
+
+/**
+ * Conceal the pitch gain of a lost subframe, ec_gain_pitch() in
+ * 3GPP TS 26.073 (ec_gains.c:306-330).
+ *
+ * @param p the context
+ * @param state state of the bad frame handling state machine
+ */
+static void ec_gain_pitch(AMRContext *p, int state)
+{
+ /* pdown[] from 3GPP TS 26.073, kept as exact fractions of 32768 */
+ static const float pdown[7] = {
+ 32767.0f / 32768.0f, 32112.0f / 32768.0f, 32112.0f / 32768.0f,
+ 26214.0f / 32768.0f, 9830.0f / 32768.0f, 6553.0f / 32768.0f,
+ 6553.0f / 32768.0f,
+ };
+ float gain = ec_gain_median(p->ec_gain_pbuf);
+
+ if (gain > p->ec_gain_past_pit)
+ gain = p->ec_gain_past_pit;
+ p->pitch_gain[4] = gain * pdown[state];
+}
+
+/**
+ * Update the pitch gain concealment state, ec_gain_pitch_update() in
+ * 3GPP TS 26.073 (ec_gains.c:343-379).
+ *
+ * @param p the context
+ * @param bfi the current frame is unusable
+ */
+static void ec_gain_pitch_update(AMRContext *p, int bfi)
+{
+ int i;
+
+ if (!bfi) {
+ /* Limit the gain of the first usable subframe after an unusable
+ frame to the last usable gain (3GPP TS 26.091 clause 7.2.2). */
+ if (p->prev_bf && p->pitch_gain[4] > p->ec_gain_prev_pit)
+ p->pitch_gain[4] = p->ec_gain_prev_pit;
+ p->ec_gain_prev_pit = p->pitch_gain[4];
+ }
+
+ p->ec_gain_past_pit = FFMIN(p->pitch_gain[4], 1.0f);
+
+ for (i = 1; i < 5; i++)
+ p->ec_gain_pbuf[i - 1] = p->ec_gain_pbuf[i];
+ p->ec_gain_pbuf[4] = p->ec_gain_past_pit;
+}
+
+/**
+ * Advance the gain predictor with the average of the four past energies,
+ * gc_pred_average_limited() followed by gc_pred_update(), as ec_gain_code()
+ * does in 3GPP TS 26.073 when the energies of the frame are not available.
+ *
+ * @param p the context
+ */
+static void ec_gain_code_predictor_update(AMRContext *p)
+{
+ float avg = 0.0f;
+ int i;
+
+ for (i = 0; i < 4; i++)
+ avg += p->prediction_error[i];
+ avg = FFMAX(avg * 0.25f, MIN_ENERGY);
+
+ memmove(&p->prediction_error[0], &p->prediction_error[1],
+ 3 * sizeof(p->prediction_error[0]));
+ p->prediction_error[3] = avg;
+}
+
+/**
+ * Conceal the codebook gain of a lost subframe, ec_gain_code() in
+ * 3GPP TS 26.073 (ec_gains.c:137-171).
+ *
+ * @param p the context
+ * @param state state of the bad frame handling state machine
+ */
+static void ec_gain_code(AMRContext *p, int state)
+{
+ /* cdown[] from 3GPP TS 26.073, kept as exact fractions of 32768 */
+ static const float cdown[7] = {
+ 32767.0f / 32768.0f, 32112.0f / 32768.0f, 32112.0f / 32768.0f,
+ 32112.0f / 32768.0f, 32112.0f / 32768.0f, 32112.0f / 32768.0f,
+ 22937.0f / 32768.0f,
+ };
+ float gain = ec_gain_median(p->ec_gain_gbuf);
+
+ if (gain > p->ec_gain_past_code)
+ gain = p->ec_gain_past_code;
+ p->fixed_gain[4] = gain * cdown[state];
+}
+
+/**
+ * Update the codebook gain concealment state, ec_gain_code_update() in
+ * 3GPP TS 26.073 (ec_gains.c:184-219).
+ *
+ * @param p the context
+ * @param bfi the current frame is unusable
+ */
+static void ec_gain_code_update(AMRContext *p, int bfi)
+{
+ int i;
+
+ if (!bfi) {
+ if (p->prev_bf && p->fixed_gain[4] > p->ec_gain_prev_code)
+ p->fixed_gain[4] = p->ec_gain_prev_code;
+ p->ec_gain_prev_code = p->fixed_gain[4];
+ }
+
+ p->ec_gain_past_code = p->fixed_gain[4];
+
+ for (i = 1; i < 5; i++)
+ p->ec_gain_gbuf[i - 1] = p->ec_gain_gbuf[i];
+ p->ec_gain_gbuf[4] = p->fixed_gain[4];
+}
+
+/// @}
+
+
/// @name AMR preprocessing functions
/// @{
@@ -1402,7 +1655,12 @@ static void cn_update_from_sid(AMRContext *p, const int *ind)
}
}
- /* Initialize the gain predictor using this SID frame's energy. */
+ /* Initialize the gain predictor using this SID frame's energy.
+
+ Note: this decoder's log2-vs-dB gain-predictor modelling (pred_init)
+ differs from the reference's fixed-point ma_pred_init (dtx_dec.c:446)
+ on the first subframe after DTX; ma_pred_init corresponds to TS 26.092
+ clause 6.2's "predictor memories are initialized". */
pred_init = FFMIN(p->cn_log_en - 9000.0f / 1024.0f, 0.0f);
pred_init = FFMAX(pred_init, MIN_ENERGY);
for (i = 0; i < 4; i++)
@@ -1446,14 +1704,14 @@ static void cn_synthesize(AMRContext *p, const float *lpc_variab, float level,
p->beta = 0.0f;
p->pitch_lag_int = 40;
- /* The reference decoder resets the error concealment states with every
- frame of a silent period, so the gains of the speech frame starting
- the next speech burst are limited to zero as well (3GPP TS 26.091
- clause 7.2.2). It also marks the frame as unusable when the comfort
- noise had been muted. */
- p->prev_pitch_gain = 0.0f;
- p->prev_fixed_gain = 0.0f;
- p->prev_bf = new_state == DTX_MUTE;
+ /* The reference decoder resets the bad frame handling and error
+ concealment states with every frame of a silent period
+ (Decoder_amr_reset() and ec_gain_*_reset() in 3GPP TS 26.073), so the
+ gains of the speech frame starting the next speech burst are limited
+ accordingly (3GPP TS 26.091 clause 7.2.2). It also marks the frame as
+ unusable when the comfort noise had been muted. */
+ ec_reset(p);
+ p->prev_bf = new_state == DTX_MUTE;
/* Decoder_amr_reset(MRDTX) clears the fixed-gain history and the
phase-dispersion state. dtx_dec() then restarts smoothing with
@@ -1595,6 +1853,39 @@ static void cn_frame(AMRContext *p, int new_state, const int *ind, int mode,
p->dtx_state = new_state;
}
+/**
+ * Fill the frame parameters of a lost speech frame with pseudo-random
+ * values, see build_CN_param() in 3GPP TS 26.073 (b_cn_cod.c:145-162).
+ * Only the fields the concealment does not itself overwrite are actually
+ * consumed (notably the fixed codebook pulses). The field order follows
+ * the AMRNBFrame order table, not the reference's bitstream parameter
+ * order, so this is deliberately not bit-exact with the reference -- the
+ * decoder is float and not bit-exact either (see the file header).
+ *
+ * @param p the context
+ * @param seed state of the pseudo-random generator
+ * @param mode speech mode of the lost frame
+ */
+static void build_bad_frame_params(AMRContext *p, uint32_t *seed, enum Mode mode)
+{
+ const uint8_t *ord = amr_unpacking_bitmaps_per_mode[mode];
+ const int16_t *win;
+
+ /* reference LCG: seed = (seed * 31821 + 13849) mod 2^16 */
+ *seed = (*seed * 31821U + 13849U) & 0xFFFFU;
+ win = &window_200_40[*seed & 0x7F];
+
+ while (*ord) {
+ int field_size = *ord++;
+ int field = *ord++; /* byte offset into AMRNBFrame */
+
+ ((uint16_t *)&p->frame)[field >> 1] =
+ *win++ & ((1 << field_size) - 1);
+
+ ord += field_size; /* skip the bit-position list */
+ }
+}
+
static int amrnb_decode_frame(AVCodecContext *avctx, AVFrame *frame,
int *got_frame_ptr, AVPacket *avpkt)
{
@@ -1623,7 +1914,8 @@ static int amrnb_decode_frame(AVCodecContext *avctx, AVFrame *frame,
return AVERROR_INVALIDDATA;
int mode = buf[0] >> 3 & 0x0F;
- int channel_size, rx_type, new_state;
+ int channel_size, rx_type, new_state, bfi;
+ int prev_dtx_state = p->dtx_state; // DTX state before this frame
int sid_ind[5] = { 0 };
int sid_mode = 0;
float synth[AMR_BLOCK_SIZE]; // synthesis before post-filtering
@@ -1702,23 +1994,41 @@ static int amrnb_decode_frame(AVCodecContext *avctx, AVFrame *frame,
continue;
}
- p->dtx_state = new_state;
+ /* A speech frame that is either flagged as bad or carries no data
+ at all cannot be decoded: its parameters are concealed instead
+ (3GPP TS 26.091 clause 7). */
+ bfi = rx_type == RX_SPEECH_BAD || rx_type == RX_NO_DATA;
- if (rx_type == RX_NO_DATA) {
- /* Outside of a silent period a frame carrying no data still
- produces a full frame of samples, keeping the frame count and
- the timing of the decoded stream correct. */
- memset(buf_out, 0, AMR_BLOCK_SIZE * sizeof(*buf_out));
- p->prev_bf = 1;
- buf += channel_size;
- buf_size -= channel_size;
- continue;
+ /* Bad frame handling state machine, 3GPP TS 26.073
+ (dec_amr.c:357-398). The state saturates at 6 and, once it has
+ reached 6, only decays to 5 on a good frame. */
+ if (bfi)
+ p->bfh_state = FFMIN(p->bfh_state + 1, 6);
+ else
+ p->bfh_state = p->bfh_state == 6 ? 5 : 0;
+
+ /* The first speech frame after a comfort noise period is seeded with
+ the DTX state of the frame before it: prev_dtx_state was saved
+ above, before this frame could overwrite dtx_state. */
+ if (prev_dtx_state == DTX) {
+ p->bfh_state = 5;
+ p->prev_bf = 0;
+ } else if (prev_dtx_state == DTX_MUTE) {
+ p->bfh_state = 5;
+ p->prev_bf = 1;
}
+ /* A frame carrying no data has no parameters of its own, fill them
+ with pseudo-random values (3GPP TS 26.073 build_CN_param()). */
+ if (rx_type == RX_NO_DATA)
+ build_bad_frame_params(p, &p->no_data_seed, p->cur_frame_mode);
+
+ p->dtx_state = new_state;
+
if (p->cur_frame_mode == MODE_12k2) {
- lsf2lsp_5(p);
+ lsf2lsp_5(p, bfi);
} else
- lsf2lsp_3(p);
+ lsf2lsp_3(p, bfi);
for (i = 0; i < 4; i++)
ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5);
@@ -1726,7 +2036,7 @@ static int amrnb_decode_frame(AVCodecContext *avctx, AVFrame *frame,
for (subframe = 0; subframe < 4; subframe++) {
const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe];
- decode_pitch_vector(p, amr_subframe, subframe);
+ decode_pitch_vector(p, amr_subframe, subframe, bfi);
decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses,
p->cur_frame_mode, subframe);
@@ -1735,8 +2045,16 @@ static int amrnb_decode_frame(AVCodecContext *avctx, AVFrame *frame,
// (section 6.1.2), but the fixed vector calculation uses
// pitch sharpening based on the on the pitch gain (section 6.1.3).
// So the correct order is: pitch gain, pitch sharpening, fixed gain.
- decode_gains(p, amr_subframe, p->cur_frame_mode, subframe,
- &fixed_gain_factor);
+ if (bfi)
+ ec_gain_pitch(p, p->bfh_state);
+ else
+ decode_gains(p, amr_subframe, p->cur_frame_mode, subframe,
+ &fixed_gain_factor);
+
+ /* Keep the concealment history and limit the gain of the first
+ usable subframe after an unusable one (3GPP TS 26.073
+ ec_gain_pitch_update(), 3GPP TS 26.091 clause 7.2.2). */
+ ec_gain_pitch_update(p, bfi);
pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse);
@@ -1747,24 +2065,20 @@ static int amrnb_decode_frame(AVCodecContext *avctx, AVFrame *frame,
ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0,
AMR_SUBFRAME_SIZE);
- p->fixed_gain[4] =
- ff_amr_set_fixed_gain(fixed_gain_factor,
- p->celpm_ctx.dot_productf(p->fixed_vector,
- p->fixed_vector,
- AMR_SUBFRAME_SIZE) /
- AMR_SUBFRAME_SIZE,
- p->prediction_error,
- energy_mean[p->cur_frame_mode], energy_pred_fac);
-
- /* 3GPP TS 26.091 clause 7.2.2: the gains of the first usable frame
- after an unusable one are limited to the ones of the last usable
- frame, so that the speech does not start at full level */
- if (p->prev_bf) {
- p->pitch_gain[4] = FFMIN(p->pitch_gain[4], p->prev_pitch_gain);
- p->fixed_gain[4] = FFMIN(p->fixed_gain[4], p->prev_fixed_gain);
+ if (bfi) {
+ ec_gain_code(p, p->bfh_state);
+ ec_gain_code_predictor_update(p);
+ } else {
+ p->fixed_gain[4] =
+ ff_amr_set_fixed_gain(fixed_gain_factor,
+ p->celpm_ctx.dot_productf(p->fixed_vector,
+ p->fixed_vector,
+ AMR_SUBFRAME_SIZE) /
+ AMR_SUBFRAME_SIZE,
+ p->prediction_error,
+ energy_mean[p->cur_frame_mode], energy_pred_fac);
}
- p->prev_pitch_gain = p->pitch_gain[4];
- p->prev_fixed_gain = p->fixed_gain[4];
+ ec_gain_code_update(p, bfi);
// The excitation feedback is calculated without any processing such
// as fixed gain smoothing. This isn't mentioned in the specification.
@@ -1815,7 +2129,7 @@ static int amrnb_decode_frame(AVCodecContext *avctx, AVFrame *frame,
highpass_gain * AMR_SAMPLE_SCALE,
p->high_pass_mem, AMR_BLOCK_SIZE);
- p->prev_bf = 0;
+ p->prev_bf = bfi;
/* keep the noise parameters of this speech frame for the next SID */
dtx_activity_update(p, synth);
diff --git a/tests/fate/amrnb.mak b/tests/fate/amrnb.mak
index 7af5b32f14..7790832f1e 100644
--- a/tests/fate/amrnb.mak
+++ b/tests/fate/amrnb.mak
@@ -62,13 +62,13 @@ fate-amrnb-state-resume-update: REF = RMS level dB: -53\.
FATE_AMRNB_DTX-$(call ALLYES, $(AMRNB_DTX_DEPS)) += fate-amrnb-nodata-ft15-q1 fate-amrnb-nodata-ft15-q0 fate-amrnb-nodata-reserved-q1
fate-amrnb-nodata-ft15-q1: CMD = ffmpeg -i $(TARGET_PATH)/tests/data/fate/amrnb-nodata-cases.amr -af "atrim=start_sample=1280:end_sample=1440,aresample=osf=flt,astats=metadata=0:reset=0" -c:a pcm_f32le -f null -
fate-amrnb-nodata-ft15-q1: CMP = grep
-fate-amrnb-nodata-ft15-q1: REF = RMS level dB: -inf
+fate-amrnb-nodata-ft15-q1: REF = RMS level dB: -52\.
fate-amrnb-nodata-ft15-q0: CMD = ffmpeg -i $(TARGET_PATH)/tests/data/fate/amrnb-nodata-cases.amr -af "atrim=start_sample=2720:end_sample=2880,aresample=osf=flt,astats=metadata=0:reset=0" -c:a pcm_f32le -f null -
fate-amrnb-nodata-ft15-q0: CMP = grep
-fate-amrnb-nodata-ft15-q0: REF = RMS level dB: -inf
+fate-amrnb-nodata-ft15-q0: REF = RMS level dB: -53\.
fate-amrnb-nodata-reserved-q1: CMD = ffmpeg -i $(TARGET_PATH)/tests/data/fate/amrnb-nodata-cases.amr -af "atrim=start_sample=4160:end_sample=4320,aresample=osf=flt,astats=metadata=0:reset=0" -c:a pcm_f32le -f null -
fate-amrnb-nodata-reserved-q1: CMP = grep
-fate-amrnb-nodata-reserved-q1: REF = RMS level dB: -inf
+fate-amrnb-nodata-reserved-q1: REF = RMS level dB: -52\.
# A SID frame arriving after a lost frame must still update the comfort
# noise. Frame 10 is the first comfort-noise frame after that SID.