FFmpeg
Functions
acelp_pitch_delay.c File Reference
#include "libavutil/common.h"
#include "libavutil/ffmath.h"
#include "libavutil/float_dsp.h"
#include "libavutil/mathematics.h"
#include "avcodec.h"
#include "acelp_pitch_delay.h"
#include "celp_math.h"
#include "audiodsp.h"

Go to the source code of this file.

Functions

void ff_acelp_update_past_gain (int16_t *quant_energy, int gain_corr_factor, int log2_ma_pred_order, int erasure)
 Update past quantized energies. More...
 
int16_t ff_acelp_decode_gain_code (AudioDSPContext *adsp, int gain_corr_factor, const int16_t *fc_v, int mr_energy, const int16_t *quant_energy, const int16_t *ma_prediction_coeff, int subframe_size, int ma_pred_order)
 Decode the adaptive codebook gain and add correction (4.1.5 and 3.9.1 of G.729). More...
 
float ff_amr_set_fixed_gain (float fixed_gain_factor, float fixed_mean_energy, float *prediction_error, float energy_mean, const float *pred_table)
 Calculate fixed gain (part of section 6.1.3 of AMR spec) More...
 
void ff_decode_pitch_lag (int *lag_int, int *lag_frac, int pitch_index, const int prev_lag_int, const int subframe, int third_as_first, int resolution)
 Decode the adaptive codebook index to the integer and fractional parts of the pitch lag for one subframe at 1/3 fractional precision. More...
 

Function Documentation

◆ ff_acelp_update_past_gain()

void ff_acelp_update_past_gain ( int16_t *  quant_energy,
int  gain_corr_factor,
int  log2_ma_pred_order,
int  erasure 
)

Update past quantized energies.

Parameters
[in,out]quant_energypast quantized energies (5.10)
gain_corr_factorgain correction factor
log2_ma_pred_orderlog2() of MA prediction order
erasureframe erasure flag

If frame erasure flag is not equal to zero, memory is updated with averaged energy, attenuated by 4dB: max(avg(quant_energy[i])-4, -14), i=0,ma_pred_order

In normal mode memory is updated with Er - Ep = 20 * log10(gain_corr_factor)

Remarks
The routine is used in G.729 and AMR (all modes).

Definition at line 32 of file acelp_pitch_delay.c.

Referenced by decode_frame().

◆ ff_acelp_decode_gain_code()

int16_t ff_acelp_decode_gain_code ( AudioDSPContext adsp,
int  gain_corr_factor,
const int16_t *  fc_v,
int  mr_energy,
const int16_t *  quant_energy,
const int16_t *  ma_prediction_coeff,
int  subframe_size,
int  max_pred_order 
)

Decode the adaptive codebook gain and add correction (4.1.5 and 3.9.1 of G.729).

Parameters
adspinitialized audio DSP context
gain_corr_factorgain correction factor (2.13)
fc_vfixed-codebook vector (2.13)
mr_energymean innovation energy and fixed-point correction (7.13)
[in,out]quant_energypast quantized energies (5.10)
subframe_sizelength of subframe
Returns
quantized fixed-codebook gain (14.1)

The routine implements equations 69, 66 and 71 of the G.729 specification (3.9.1)

Em - mean innovation energy (dB, constant, depends on decoding algorithm) Ep - mean-removed predicted energy (dB) Er - mean-removed innovation energy (dB) Ei - mean energy of the fixed-codebook contribution (dB) N - subframe_size M - MA (Moving Average) prediction order gc - fixed-codebook gain gc_p - predicted fixed-codebook gain

Fixed codebook gain is computed using predicted gain gc_p and correction factor gain_corr_factor as shown below:

gc = gc_p * gain_corr_factor

The predicted fixed codebook gain gc_p is found by predicting the energy of the fixed-codebook contribution from the energy of previous fixed-codebook contributions.

mean = 1/N * sum(i,0,N){ fc_v[i] * fc_v[i] }

Ei = 10log(mean)

Er = 10log(1/N * gc^2 * mean) - Em = 20log(gc) + Ei - Em

Replacing Er with Ep and gc with gc_p we will receive:

Ep = 10log(1/N * gc_p^2 * mean) - Em = 20log(gc_p) + Ei - Em

and from above:

gc_p = 10^((Ep - Ei + Em) / 20)

Ep is predicted using past energies and prediction coefficients:

Ep = sum(i,0,M){ ma_prediction_coeff[i] * quant_energy[i] }

gc_p in fixed-point arithmetic is calculated as following:

mean = 1/N * sum(i,0,N){ (fc_v[i] / 2^13) * (fc_v[i] / 2^13) } =
= 1/N * sum(i,0,N) { fc_v[i] * fc_v[i] } / 2^26

Ei = 10log(mean) = -10log(N) - 10log(2^26) +
+ 10log(sum(i,0,N) { fc_v[i] * fc_v[i] })

Ep - Ei + Em = Ep + Em + 10log(N) + 10log(2^26) -
- 10log(sum(i,0,N) { fc_v[i] * fc_v[i] }) =
= Ep + mr_energy - 10log(sum(i,0,N) { fc_v[i] * fc_v[i] })

gc_p = 10 ^ ((Ep - Ei + Em) / 20) =
= 2 ^ (3.3219 * (Ep - Ei + Em) / 20) = 2 ^ (0.166 * (Ep - Ei + Em))

where

mr_energy = Em + 10log(N) + 10log(2^26)
Remarks
The routine is used in G.729 and AMR (all modes).

Definition at line 53 of file acelp_pitch_delay.c.

Referenced by decode_frame().

◆ ff_amr_set_fixed_gain()

float ff_amr_set_fixed_gain ( float  fixed_gain_factor,
float  fixed_mean_energy,
float prediction_error,
float  energy_mean,
const float pred_table 
)

Calculate fixed gain (part of section 6.1.3 of AMR spec)

Parameters
fixed_gain_factorgain correction factor
fixed_mean_energymean decoded algebraic codebook vector energy
prediction_errorvector of the quantified predictor errors of the four previous subframes. It is updated by this function.
energy_meandesired mean innovation energy
pred_tabletable of four moving average coefficients

Definition at line 86 of file acelp_pitch_delay.c.

Referenced by amrnb_decode_frame(), amrwb_decode_frame(), and decode_frame().

◆ ff_decode_pitch_lag()

void ff_decode_pitch_lag ( int lag_int,
int lag_frac,
int  pitch_index,
const int  prev_lag_int,
const int  subframe,
int  third_as_first,
int  resolution 
)

Decode the adaptive codebook index to the integer and fractional parts of the pitch lag for one subframe at 1/3 fractional precision.

The choice of pitch lag is described in 3GPP TS 26.090 section 5.6.1.

Parameters
lag_intinteger part of pitch lag of the current subframe
lag_fracfractional part of pitch lag of the current subframe
pitch_indexparsed adaptive codebook (pitch) index
prev_lag_intinteger part of pitch lag for the previous subframe
subframecurrent subframe number
third_as_firsttreat the third frame the same way as the first

Definition at line 107 of file acelp_pitch_delay.c.

Referenced by decode_frame(), and decode_pitch_vector().