libavcodec/amrnbdec.c File Reference

AMR narrowband decoder. More...

#include <string.h>
#include <math.h>
#include "avcodec.h"
#include "get_bits.h"
#include "libavutil/common.h"
#include "celp_math.h"
#include "celp_filters.h"
#include "acelp_filters.h"
#include "acelp_vectors.h"
#include "acelp_pitch_delay.h"
#include "lsp.h"
#include "amrnbdata.h"

Go to the source code of this file.

Data Structures

struct AMRContext

Defines

#define AMR_BLOCK_SIZE   160

samples per frame

#define AMR_SAMPLE_BOUND   32768.0

threshold for synthesis overflow

#define AMR_SAMPLE_SCALE   (2.0 / 32768.0)

Scale from constructed speech to [-1,1].

#define PRED_FAC_MODE_12k2   0.65

Prediction factor for 12.2kbit/s mode.

#define LSF_R_FAC   (8000.0 / 32768.0)

LSF residual tables to Hertz.

#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.

#define MIN_ENERGY   -14.0

Initial energy in dB.

#define SHARP_MAX   0.79449462890625

Maximum sharpening factor.

#define AMR_TILT_RESPONSE   22

Number of impulse response coefficients used for tilt factor.

#define AMR_TILT_GAMMA_T   0.8

Tilt factor = 1st reflection coefficient * gamma_t.

#define AMR_AGC_ALPHA   0.9

Adaptive gain control factor used in post-filter.

Functions

static void weighted_vector_sumd (double *out, const double *in_a, const double *in_b, double weight_coeff_a, double weight_coeff_b, int length)

Double version of ff_weighted_vector_sumf().

static av_cold int amrnb_decode_init (AVCodecContext *avctx)

static enum Mode unpack_bitstream (AMRContext *p, const uint8_t *buf, int buf_size)

Unpack an RFC4867 speech frame into the AMR frame mode and parameters.

static void lsf2lsp (const float *lsf, double *lsp)

Convert an lsf vector into an lsp vector.

static void interpolate_lsf (float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)

Interpolate the LSF vector (used for fixed gain smoothing).

static void lsf2lsp_for_mode12k2 (AMRContext *p, double lsp[LP_FILTER_ORDER], const float lsf_no_r[LP_FILTER_ORDER], const int16_t *lsf_quantizer[5], const int quantizer_offset, const int sign, const int update)

Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector.

static void lsf2lsp_5 (AMRContext *p)

Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors.

static void lsf2lsp_3 (AMRContext *p)

Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector.

static void decode_pitch_lag_1_6 (int *lag_int, int *lag_frac, int pitch_index, const int prev_lag_int, const int subframe)

Like ff_decode_pitch_lag(), but with 1/6 resolution.

static void decode_pitch_vector (AMRContext *p, const AMRNBSubframe *amr_subframe, const int subframe)

static void decode_10bit_pulse (int code, int pulse_position[8], int i1, int i2, int i3)

Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame.

static void decode_8_pulses_31bits (const int16_t *fixed_index, AMRFixed *fixed_sparse)

Decode the algebraic codebook index to pulse positions and signs and construct the algebraic codebook vector for MODE_10k2.

static void decode_fixed_sparse (AMRFixed *fixed_sparse, const uint16_t *pulses, const enum Mode mode, const int subframe)

Decode the algebraic codebook index to pulse positions and signs, then construct the algebraic codebook vector.

static void pitch_sharpening (AMRContext *p, int subframe, enum Mode mode, AMRFixed *fixed_sparse)

Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2).

static float fixed_gain_smooth (AMRContext *p, const float *lsf, const float *lsf_avg, const enum Mode mode)

fixed gain smoothing Note that where the spec specifies the "spectrum in the q domain" in section 6.1.4, in fact frequencies should be used.

static void decode_gains (AMRContext *p, const AMRNBSubframe *amr_subframe, const enum Mode mode, const int subframe, float *fixed_gain_factor)

Decode pitch gain and fixed gain factor (part of section 6.1.3).

static void apply_ir_filter (float *out, const AMRFixed *in, const float *filter)

Circularly convolve a sparse fixed vector with a phase dispersion impulse response filter (D.6.2 of G.729 and 6.1.5 of AMR).

static const float * anti_sparseness (AMRContext *p, AMRFixed *fixed_sparse, const float *fixed_vector, float fixed_gain, float *out)

Reduce fixed vector sparseness by smoothing with one of three IR filters.

static int synthesis (AMRContext *p, float *lpc, float fixed_gain, const float *fixed_vector, float *samples, uint8_t overflow)

Conduct 10th order linear predictive coding synthesis.

static void update_state (AMRContext *p)

Update buffers and history at the end of decoding a subframe.

static float tilt_factor (float *lpc_n, float *lpc_d)

Get the tilt factor of a formant filter from its transfer function.

static void postfilter (AMRContext *p, float *lpc, float *buf_out)

Perform adaptive post-filtering to enhance the quality of the speech.

static int amrnb_decode_frame (AVCodecContext *avctx, void *data, int *data_size, AVPacket *avpkt)

Variables

AVCodec amrnb_decoder

Detailed Description

AMR narrowband decoder.

This decoder uses floats for simplicity and so is not bit-exact. One difference is that differences in phase can accumulate. The test sequences in 3GPP TS 26.074 can still be useful.

Comparing this file's output to the output of the ref decoder gives a PSNR of 30 to 80. Plotting the output samples shows a difference in phase in some areas.

Comparing both decoders against their input, this decoder gives a similar PSNR. If the test sequence homing frames are removed (this decoder does not detect them), the PSNR is at least as good as the reference on 140 out of 169 tests.

Definition in file amrnbdec.c.

Define Documentation

#define AMR_AGC_ALPHA 0.9

Adaptive gain control factor used in post-filter.

Definition at line 94 of file amrnbdec.c.

Referenced by postfilter().

#define AMR_BLOCK_SIZE 160

samples per frame

Definition at line 58 of file amrnbdec.c.

Referenced by amrnb_decode_frame().

#define AMR_SAMPLE_BOUND 32768.0

threshold for synthesis overflow

Definition at line 59 of file amrnbdec.c.

Referenced by synthesis().

#define AMR_SAMPLE_SCALE (2.0 / 32768.0)

Scale from constructed speech to [-1,1].

AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but upscales by two (section 6.2.2).

Fundamentally, this scale is determined by energy_mean through the fixed vector contribution to the excitation vector.

Definition at line 70 of file amrnbdec.c.

Referenced by amrnb_decode_frame().

#define AMR_TILT_GAMMA_T 0.8

Tilt factor = 1st reflection coefficient * gamma_t.

Definition at line 92 of file amrnbdec.c.

Referenced by tilt_factor().

#define AMR_TILT_RESPONSE 22

Number of impulse response coefficients used for tilt factor.

Definition at line 90 of file amrnbdec.c.

Referenced by tilt_factor().

#define LSF_R_FAC (8000.0 / 32768.0)

LSF residual tables to Hertz.

Definition at line 75 of file amrnbdec.c.

Referenced by lsf2lsp_3(), lsf2lsp_5(), and lsf2lsp_for_mode12k2().

#define MIN_ENERGY -14.0

Initial energy in dB.

Also used for bad frames (unimplemented).

Definition at line 80 of file amrnbdec.c.

Referenced by amrnb_decode_init().

#define MIN_LSF_SPACING (50.0488 / 8000.0)

Ensures stability of LPC filter.

Definition at line 76 of file amrnbdec.c.

Referenced by lsf2lsp_3(), and lsf2lsp_for_mode12k2().

#define PITCH_LAG_MIN_MODE_12k2 18

Lower bound on decoded lag search in 12.2kbit/s mode.

Definition at line 77 of file amrnbdec.c.

Referenced by decode_pitch_lag_1_6().

#define PRED_FAC_MODE_12k2 0.65

Prediction factor for 12.2kbit/s mode.

Definition at line 73 of file amrnbdec.c.

Referenced by lsf2lsp_5().

#define SHARP_MAX 0.79449462890625

Maximum sharpening factor.

The specification says 0.8, which should be 13107, but the reference C code uses 13017 instead. (Amusingly the same applies to SHARP_MAX in g729dec.c.)

Definition at line 87 of file amrnbdec.c.

Referenced by decode_frame(), pitch_sharpening(), and synthesis().

Function Documentation

static int amrnb_decode_frame	(	AVCodecContext *	avctx,
		void *	data,
		int *	data_size,
		AVPacket *	avpkt
	)			`[static]`

Definition at line 950 of file amrnbdec.c.

static av_cold int amrnb_decode_init ( AVCodecContext * avctx ) [static]

Definition at line 151 of file amrnbdec.c.

static enum Mode unpack_bitstream	(	AMRContext *	p,
		const uint8_t *	buf,
		int	buf_size
	)			`[static]`

Unpack an RFC4867 speech frame into the AMR frame mode and parameters.

The order of speech bits is specified by 3GPP TS 26.101.

Parameters:

	p	the context
	buf	pointer to the input buffer
	buf_size	size of the input buffer

Returns:: the frame mode

Definition at line 184 of file amrnbdec.c.

Referenced by amrnb_decode_frame().

static void weighted_vector_sumd	(	double *	out,
		const double *	in_a,
		const double *	in_b,
		double	weight_coeff_a,
		double	weight_coeff_b,
		int	length
	)			`[static]`

Double version of ff_weighted_vector_sumf().

Definition at line 140 of file amrnbdec.c.

Referenced by lsf2lsp_5().

Variable Documentation

AVCodec amrnb_decoder

Initial value:

 {
    .name           = "amrnb",
    .type           = AVMEDIA_TYPE_AUDIO,
    .id             = CODEC_ID_AMR_NB,
    .priv_data_size = sizeof(AMRContext),
    .init           = amrnb_decode_init,
    .decode         = amrnb_decode_frame,
    .long_name      = NULL_IF_CONFIG_SMALL("Adaptive Multi-Rate NarrowBand"),
    .sample_fmts    = (enum SampleFormat[]){SAMPLE_FMT_FLT,SAMPLE_FMT_NONE},
}

Definition at line 1067 of file amrnbdec.c.


Data Structures
struct	AMRContext
Defines
#define	AMR_BLOCK_SIZE 160
	samples per frame
#define	AMR_SAMPLE_BOUND 32768.0
	threshold for synthesis overflow
#define	AMR_SAMPLE_SCALE (2.0 / 32768.0)
	Scale from constructed speech to [-1,1].
#define	PRED_FAC_MODE_12k2 0.65
	Prediction factor for 12.2kbit/s mode.
#define	LSF_R_FAC (8000.0 / 32768.0)
	LSF residual tables to Hertz.
#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.
#define	MIN_ENERGY -14.0
	Initial energy in dB.
#define	SHARP_MAX 0.79449462890625
	Maximum sharpening factor.
#define	AMR_TILT_RESPONSE 22
	Number of impulse response coefficients used for tilt factor.
#define	AMR_TILT_GAMMA_T 0.8
	Tilt factor = 1st reflection coefficient * gamma_t.
#define	AMR_AGC_ALPHA 0.9
	Adaptive gain control factor used in post-filter.
Functions
static void	weighted_vector_sumd (double out, const double in_a, const double *in_b, double weight_coeff_a, double weight_coeff_b, int length)
	Double version of ff_weighted_vector_sumf().
static av_cold int	amrnb_decode_init (AVCodecContext *avctx)
static enum Mode	unpack_bitstream (AMRContext p, const uint8_t buf, int buf_size)
	Unpack an RFC4867 speech frame into the AMR frame mode and parameters.
static void	lsf2lsp (const float lsf, double lsp)
	Convert an lsf vector into an lsp vector.
static void	interpolate_lsf (float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
	Interpolate the LSF vector (used for fixed gain smoothing).
static void	lsf2lsp_for_mode12k2 (AMRContext p, double lsp[LP_FILTER_ORDER], const float lsf_no_r[LP_FILTER_ORDER], const int16_t lsf_quantizer[5], const int quantizer_offset, const int sign, const int update)
	Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector.
static void	lsf2lsp_5 (AMRContext *p)
	Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors.
static void	lsf2lsp_3 (AMRContext *p)
	Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector.
static void	decode_pitch_lag_1_6 (int lag_int, int lag_frac, int pitch_index, const int prev_lag_int, const int subframe)
	Like ff_decode_pitch_lag(), but with 1/6 resolution.
static void	decode_pitch_vector (AMRContext p, const AMRNBSubframe amr_subframe, const int subframe)
static void	decode_10bit_pulse (int code, int pulse_position[8], int i1, int i2, int i3)
	Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame.
static void	decode_8_pulses_31bits (const int16_t fixed_index, AMRFixed fixed_sparse)
	Decode the algebraic codebook index to pulse positions and signs and construct the algebraic codebook vector for MODE_10k2.
static void	decode_fixed_sparse (AMRFixed fixed_sparse, const uint16_t pulses, const enum Mode mode, const int subframe)
	Decode the algebraic codebook index to pulse positions and signs, then construct the algebraic codebook vector.
static void	pitch_sharpening (AMRContext p, int subframe, enum Mode mode, AMRFixed fixed_sparse)
	Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2).
static float	fixed_gain_smooth (AMRContext p, const float lsf, const float *lsf_avg, const enum Mode mode)
	fixed gain smoothing Note that where the spec specifies the "spectrum in the q domain" in section 6.1.4, in fact frequencies should be used.
static void	decode_gains (AMRContext p, const AMRNBSubframe amr_subframe, const enum Mode mode, const int subframe, float *fixed_gain_factor)
	Decode pitch gain and fixed gain factor (part of section 6.1.3).
static void	apply_ir_filter (float out, const AMRFixed in, const float *filter)
	Circularly convolve a sparse fixed vector with a phase dispersion impulse response filter (D.6.2 of G.729 and 6.1.5 of AMR).
static const float *	anti_sparseness (AMRContext p, AMRFixed fixed_sparse, const float fixed_vector, float fixed_gain, float out)
	Reduce fixed vector sparseness by smoothing with one of three IR filters.
static int	synthesis (AMRContext p, float lpc, float fixed_gain, const float fixed_vector, float samples, uint8_t overflow)
	Conduct 10th order linear predictive coding synthesis.
static void	update_state (AMRContext *p)
	Update buffers and history at the end of decoding a subframe.
static float	tilt_factor (float lpc_n, float lpc_d)
	Get the tilt factor of a formant filter from its transfer function.
static void	postfilter (AMRContext p, float lpc, float *buf_out)
	Perform adaptive post-filtering to enhance the quality of the speech.
static int	amrnb_decode_frame (AVCodecContext avctx, void data, int data_size, AVPacket avpkt)
Variables
AVCodec	amrnb_decoder