4 C code structure

26.2433GPPANSI-C code for the fixed-point distributed speech recognition extended advanced front-endRelease 17TS

This clause gives an overview of the structure of the bit‑exact C code and provides an overview of the contents and organization of the C code attached to this document.

The C code has been verified on the following systems:

– Sun Microsystems workstations and GNU gcc compiler

– IBM PC compatible computers with Linux operating system and GNU gcc compiler.

ANSI‑C was selected as the programming language because portability was desirable.

4.1 Contents of the C source code

The distributed files with suffix "c" contain the source code and the files with suffix "h" are the header files.

Makefiles are provided for the platforms in which the C code has been verified (listed above).

4.2 Program execution

There are separate executables for the FrontEnd and Vector Quantization, with and without Extensions. The command line options are described below.

<> – indicates parameters for the given option for running the executable

() – indicates default parameter.

FrontEnd w/ Extension:

USAGE: bin/ExtAdvFrontEnd infile HTK_outfile pitch_outfile class_outfile [options]

OPTIONS:

-q Quiet Mode (FALSE)

-F format Input file format <NIST,HTK,RAW> (NIST)

-fs freq Sampling frequency in kHz <8,16> (8)

-swap Change input byte ordering (Native)

-noh No HTK header to output file (FALSE)

-noc0 No c0 coefficient to output feature vector (FALSE)

-nologE No logE component to output feature vector (FALSE)

-skip_header_bytes n – Skip header, first n bytes ( Only for -F RAW)

-noh, -noc0, -nologE and –skip_header_bytes are not used and should not be changed.

FrontEnd w/o Extension:

USAGE: bin/AdvFrontEnd infile HTK_outfile [options]

OPTIONS: – Same as FrontEnd w/ Extension

Vector Quantization w/ Extension:

Usage: extcoder htk_file_in pitch_file_in class_file_in bitstream_file_out pitch_file_out txt_file_out -freq x -VAD/No_VAD

htk_file_in Input mel-frequency cepstral coefficient file in HTK MFCC format.

pitch_file_in Input pitch period file.

class_file_in Input classification file.

bit_file_out Output binary bitstream.

pitch_file_out Output quantised pitch period file.

txt_file_out Vector quantiser output in text format.

-freq x Sampling frequency in kHz (8 or 16).

-VAD Use voice activity detector data. Voice activity input file must have same name as htk_file, but extension .vad

-No_VAD Do not incorporate voice activity detector information in output bitstream.

Vector Quantization w/o Extension:

Usage: coder htk_file_in bitstream_file_out txt_file_out -freq x -VAD/No_VAD

htk_file_in Input mel-frequency cepstral coefficient file in HTK MFCC format.

bit_file_out Binary output bitstream.

txt_file_out Vector quantiser output in text format.

-freq x Sampling frequency in kHz (8 or 16).

-VAD Use voice activity detector data. Voice activity input file must have same name as htk_file, but extension .vad

-No_VAD Do not incorporate voice activity detector information in output bitstream.

File extension descriptions as generated by the sample script:

.cep – Binary file containing cepstral features in HTK format. Output from the FrontEnd, input to the vector quantizer.

.pitch – Binary file containing pitch information. Output from the FrontEnd, input to the vector quantizer. Only used for Extension.

.class – Ascii file containing class information. Output from the FrontEnd, input to the vector quantizer. Only used for Extension.

.bs – Binary file containing the bitstream. Output from the vector quantizer.

.log – Log files from the different executables.

4.3 Code hierarchy

Tables 1 to 3 are call graphs that show the functions used for AFE (table 1), VQ (table 2), and Extension (table 3).

Each column represents a call level and each cell a function. The functions contain calls to the functions in rightwards neighboring cells. The time order in the call graphs is from the top downwards as the processing of a frame advances. All standard C functions: printf(), fwrite(), etc. have been omitted. Also, no basic operations (add(), L_add(), mac(), etc.) or double precision extended operations (e.g. L_Extract()) appear in the graphs.

The basic operations are not counted as extending the depth, therefore the deepest level in this software is level 7.

Table 1: AFE call structure

main()

AdvProcessInit_B()

DoNoiseSupInit_B()

DoWaveProcInit_B()

DoCompCepsInit_B()

DoPostProcInit_B()

DoVADInit_F()

Do16kProcInit_B()

QMF_FIR_Init_B()

fir_initialization_B()

DP_HP_filters_B()

BufIn32Alloc()

AdvProcessAlloc_B()

DoNoiseSupAlloc_B()

DoWaveProcAlloc_B()

DoCompCepsAlloc_B()

DoPostProcAlloc_B()

DoVADAlloc_F()

Do16kProcAlloc_B()

FlushAdvProcess_B()

DoVADFlush_F()

CvFeatInt2Float()

AdvProcessDelete_B()

DoNoiseSupDelete_B()

DoWaveProcDelete_B()

DoCompCepsDelete_B()

DoPostProcDelete_B()

DoVADDelete_B()

BufIn32Free()

DoAdvProcess_B()

Do16kProcessing_B()

DoNoiseSup_B()

Get16k_p_bufferData16k_B()

Get16k_bufData16kSize_B()

Get16k_p_BandsForCoding16k_B()

Get16k_p_CodeForBands16k_B()

Get16k_dataHP_B()

VAD_F()

Log_2()

DoSigWindowing16_F1()

DoSigWindowing16_F2()

ff4NRFix32_B()

GetL15()

GetH15()

Mult16x32()

Add_Mult16x16_16()

Sub_Mult16x16_16()

Permut()

FFTtoPSD_F()

Square24d2_B()

Square24_B()

Get16k_BFC_dec_B()

GetBandsForCoding16k_B()

PSDMean_F()

NoiseEstimation_F1()

Sqrt_2()

Sqrt16_2()

NoiseEstimation_F2()

Sqrt_2()

Sqrt16_2()

FilterCalc_F()

SpeechQVar()

FilterBank16()

SpeechQSpec()

SpeechQMel()

DoGainFact_F1()

Log_2()

DoGainFact_F2()

Log_2()

DoMelIDCT_F16()

ApplyWF()

Get16k_dec1()

Get16k_dec2()

Get16k_dec3()

DoSigWindowing16_F3()

ff4NRFix32_B()

GetL15()

GetH15()

Mult16x32()

Add_Mult16x16_16()

Sub_Mult16x16_16()

Permut()

FFTtoPSD_F()

Square24d2_B()

Square24_B()

DoMelFB_B()

CodeBands16k_B()

DoSpecSub16k_B()

Log_2()

UpDateDecal()

ApplyDecal()

DCOffsetFil_F()

Get16k_hpBandsSize_B()

Get16k_p_hpBands_B()

Get16k_p_bufferCodeForBands16k_B()

Get16k_p_CodeForBands16k_B()

Get16k_p_bufferCodeWeights_B()

Get16k_p_codeWeights_B()

Set16k_hpBands_dec_B()

DoWaveProc_B()

TeagerEng()

GetTeagerFilter()

GetMaximaPositions()

DoCompCeps_B()

CepsCompute()

Get16k_p_bufferCodeWeights_B()

Get16k_p_bufferCodeForBands16k_B()

PreEmphHamm()

ff4NB16_B()

GetBandsForDecoding16k_B()

DecodeBands16k_B()

FilterBank()

Get16k_hpBands_dec_B()

Get16k_p_hpBands_B()

MergeSSandCoded_B()

CorrectEnergy_B()

CosInv16Khz()

cosInv() (only for 8kHz)

DoPostProc_B()

DoVADProc_F()

focalpoint()

Table 2: VQ call structure

main()

quantize_and_print()

get_best_dataframe()

best_centroid()

quant_pitch_abs()

get_class_bit()

quant_pitch_diff()

get_class_bit()

mfcc_crc_encode()

pc_crc_encode()

Table 3: Extension call structure

main()

RVC_ConstructPitchRom_be()

RVC_ConstructPitchMeter_be()

Allocate_InterpolatedDft_be()

RVC_ResetPitchMeter_be()

RVC_DestructPitchRom_be()

RVC_DestructPitchMeter_be()

Deallocate_InterpolatedDft_be()

DoAdvProcess_B()

DoPitchExtract()

FilterBank()

dsr_afe_vad()

get_vm()

fnLog2()

IsLowBandNoise()

get_zcm()

pre_process()

iir_d()

iir_s()

RVC_MeasurePitch_be()

ClearPitch_be()

DirichletInterpolation_be()

IsLowLevelInput_be()

Finalize_be()

IsContinuousPitch_be()

Mpy_lw_sw()

Mpy_lw_sw()

PrepareSpectralPeaks_be()

CalcSpectrum_be()

Mpy_lw_sw()

Mpy_lw_sw_Add()

FindPeaks_be()

Prelim_ScaleDownAmpsOfHighFreqPeaks_be()

qsort_be()*

swap()

CompareIpointAmp_be()

RefineSpectralPeaks_be()

sqrt_l_fix()

Final_ScaleDownAmpsOfHighFreqPeaks_be()

Mpy_lw_sw()

FindPitchCandidates_be()

NormalizeAmplitudes_be()

CalcUtilityFunction_be()

CreatePieceWiseConstantFunction_be()

L_Extract()

Mpy_32_16()

qsort_be()*

swap()

Compare_ARRAY_OF_XPOINTS_be()

LinkArrayOfPoints_be()

AddSortedArrayOfPoints_be()

LinkArrayOfPoints_be()

ConvertLinkedListOfDiffPointsToUtilFunc_be()

FindDominantLocalMaximaInUtilityFunction_be()

Mpy_lw_sw()

UtilityFunctionAtGivenPitchFreq_be()

qsort_be()*

swap()

ComparePitchFreqAscending_be()

SelectTopPitchCandidates_be()

Mpy_lw_sw()

compute_pcorr_be()

interpolate_be()

Mpy_lw_sw()

Mpy_lw_lw()

sqrt_l_fix()

find_most_energetic_window_be()

accumulate_be()

find_most_energetic_window2_be()

Mpy_lw_sw()

SelectFinalPitch_be()

qsort_be()*

swap()

ComparePitchFreqDescending_be()

ClearPitch_be()

GOOD_ENOUGH_be()

CLOSELY_LOCATED_be()

Mpy_lw_sw()

BETTER_be()

IsContinuousPitch_be()

Mpy_lw_sw()

CalculateDoubleWindowDft_be()

classify_frame()

* qsort_be() is a recursive function

4.5 Variables, constants and tables

The data types of variables and tables used in the fixed point implementation are signed integers in 2’s complement representation, defined by:

– Word16 16 bit variable;

– Word32 32 bit variable.

4.5.1 Description of constants used in the C-code

Table 5a: Global constants for AFE

Constant

Value

Description

NS_SPEC_ORDER_16K

64

Noise suppression Array length

NS_HANGOVER_16K

15

Noise suppression hangover count

NS_MIN_SPEECH_FRAME_HANGOVER_16K

4

Noise suppression minmum speech frame hangover count

NS_ANALYSIS_WINDOW_16K

80

Noise suppression analysis window

PERC_CODED

0.7

lambda merge (empirically set constant)

LAMBDA_NSE16k

0.99

Noise estimation Lambda

NS_NB_FRAME_THRESHOLD_NSE

100

Noise suppression number of frame threshold used for NSE

LENGTH_QMF

118

QMF filter length

f24

1

multiplier for QMF filter coefficients

SHFF_H

8

shift to get higher value

L_H

16

shift to get lower value

HP16k_MEL_USED

3

Higher frequnecy band Mel used

NB_LP_BANDS_CODING

3

Lower frequency band used in coding

NE16k_FRAMES_THRESH

100

Noise estimation frames threshold

NB_TOPOSTPROC

12

Number of coefficients to postprocess

CEP_FRAME_LENGTH

200

Frame length for cepstral coefficients

CEP_NB_COEF

13

Number of cepstral coefficients (including c0)

CEP_NB_CHANNELS

23

Number of filters used for cepstral coefficients

CEP_FFT_LENGTH

256

FFT length for cepstral coefficients

FRAME_BUF_SIZE

241

Denoised Output buffer size

FRAME_SHIFT

80

WaveProcessing input frame shift

FRAME_LENGTH

200

WaveProcessing frame size

NS_SPEC_ORDER

65

Noise suppression array length (8khz)

NS_BUFFER_SIZE

180

Noise suppression past frame size

NS_FRAME_SHIFT

80

Noise suppression input frame shift

NS_HALF_FILTER_LENGTH

8

Noise suppression filter half size

NS_NB_FRAME_THRESHOLD_LTE

10

Noise suppression long term energy forgetting factor threshold (in frames)

NS_NB_FRAME_THRESHOLD_NSE

100

Noise suppression spectrum estimate forgetting factor threshold (in frames)

NS_MIN_FRAME

10

Number of frame threshold to update average energy for Nosie suppression VAD

NS_FFT_LENGTH

256

FFT length for noise suppression

WF_MEL_ORDER

25

Noise suppression Wiener filter order

SHFT_NOISE

14

shift applied to noise spectrum estimate

SHFT_FACT_MUL

14

shift applied to gain coefficient (nosie suppression gain factoriization)

IDCT_ORDER

25

Noise suppression idct order

NS_BETA

0.98

Noiseless signal suppression factor

NS_RSB_MIN

0.079432823

Minimum a priori SNR

NS_LAMBDA_NSE

0.99

Forgetting factor for noise spectrum estimate

NS_LOG_SPEC_FLOOR

-10.0

average energy minimum threshold

NS_SNR_THRESHOLD_VAD

15

SNR threshold for noise suppression VAD

NS_SNR_THRESHOLD_UPD_LTE

20

Long term energy update threshold for noise suppression VAD

NS_ENERGY_FLOOR

80

Energy Minimum threshold for noise suppression VAD

MaxPos

10

Maximum number of maxima in waveprocessing

WP_EPS

0.2

weigthing value added or substracted for waveprocessing

Table 5b: Global constants for VQ

Constant

Value

Description

MIN_PERIOD

1245184

Minimum pitch period allowed

MAX_PERIOD

9175040

Maximum pitch period allowed

NUM_MULTI_LEVELS_1

26

number of levels in pitch quantization

NUM_MULTI_LEVELS_2

24

number of levels in pitch quantization

UNVOICED_CODE

0

init value for Qpindex

Table 5c: Global constants for Extension

Constant

Value

Description

HISTORY_LEN

100

History length – past samples for pitch extraction

DOWN_SAMP_FACTOR

4

Down-sampling factor – used in computing correlation

NO_OF_DFT_POINTS

128

Number of DFT points

BREAK_POINT

12

Break point – marks the end of low frequency band

LBN_HIST_WEIGHT

32440

Low band noise history weight

LBN_CURR_WEIGHT

328

Low band noise current weight (32768 – LBN_HIST_WEIGHT)

LBN_MAX_THR

124518

Low band noise maximum threshold

LBN_LOW_ENR_LEVEL_MANT

32000

Low band noise low energy level mantissa

LBN_LOW_ENR_LEVEL_SHFT

22

Low band noise low energy level shift

RVC_OK

0

Return code for success

RVC_ERR

-1

Return code for unspecified error

RVC_ERR_NOT_ENOUGH_MEMORY

-2

Return code for not enough memory

RVC_ERR_ILLEGAL_ARGUMENT

-3

Return code for an illegal input / output argument

RVC_ERR_IO_FAILED

-4

Return code for failed input / output to a file

RVC_ERR_BAD_FILE_FORMAT

-5

Return code for a bad file header

RVC_ERR_NOT_INITIALIZED

-6

Return code for failure due to improper initialization

RVC_ERR_ILLEGAL_USAGE

-7

Return code for illegal usage of a function

RVC_ERR_NOT_ENOUGH_SAMPLES

-8

Return code for insufficient number of samples

RVC_ERR_NOT_IMPLEMENTED

-9

Return code for an unimplemented function

RVC_ERR_FAIL_OPEN_FILE

-10

Return code for failure to open a file

UB_ENRG_FRAC

59

Upper band energy fraction

ZCM_THLD

87

Zero crossing measure threshold

SQRT_ONE_HALF

0x5A82

Square root of 0.5 (0.707)

FRAME_LEN_DS

50

Frame length downsampled (200/4)

FRAME_LEN_DS_BY_2

25

Frame length downsampled divided by 2

HISTORY_LEN_DS

25

History length downsampled (100/4)

WINDOW_LENGTH

18

Window length used in computing correlation

INV_WINDOW_LENGTH

1820

Inverse of window length (1/18 = 0.05556)

NUM_CHAN

23

Number of channels or Mel-frequency bands

MIN_CH_ENRG_MANTISSA

20000

Minimum channel energy mantissa

MIN_CH_ENRG_SHIFT

25

Minimum channel energy shift

INIT_SIG_ENRG_MANTISSA

30518

Initial signal energy mantissa

INIT_SIG_ENRG_SHIFT

8

Initial signal energy shift

CE_SM_FAC

18022

Channel energy smoothing factor

CE_SM_FAC_COMPL

14746

Channel energy smoothing factor complement

CNE_SM_FAC

3277

Channel noise energy smoothing factor

CNE_SM_FAC_COMPL

29491

Channel noise energy smoothing factor complement

LO_GAMMA

22938

Low gamma value

LO_GAMMA_COMPL

9830

Low gamma value complement

HI_GAMMA

29491

High gamma value

HI_GAMMA_COMPL

3277

High gamma value complement

LO_BETA

31130

Low beta value

HI_BETA

32702

High beta value

INIT_FRAMES

10

Initial number of frames (considered to be noise frames)

SINE_START_CHAN

4

Sine start channel (for sine wave detection)

PEAK_TO_AVE_THLD

10

Peak to average threshold

DEV_THLD

1523942

Deviation threshold

HYSTER_CNT_THLD

9

Hysteresis count threshold

F_UPDATE_CNT_THLD

500

Forced update count threshold

NON_SPEECH_THLD

32

Non-speech threshold

FIX_34

24576

(short) (32768.0 * 3.0/4.0)

FIX_18

4096

(short) (32768.0 * 1.0/8.0)

FIX_INVSQRT2

-23170

1 / sqrt(2)

swTHIRD_REF_BANDWIDTH

85

One third of the reference bandwidth

swTWO_THIRDS_REF_BANDWIDTH

171

Two thirds of the reference bandwidth

MIN_ENERGY_MANTISSA

25600

Minimum energy mantissa

MIN_ENERGY_SHIFT

18

Minimum energy shift

swREF_SAMPLE_RATE_Q0

0x1F40

Reference sampling rate in Q0 format

swCLOSE_FACTOR_Q14

0x4CCD

Closeness factor in Q14 format

swFD_SCORE_THLD1_Q15

0x63D7

Frequency domain score threshold 1 in Q15 format

swFD_SCORE_THLD2_Q15

0x570A

Frequency domain score threshold 2 in Q15 format

swCORR_THLD_Q15

0x651F

Correlation threshold in Q15 format

swSUM_THLD_Q14

0x6667

Sum threshold in Q14 format

lwCRIT0_OFFSET_Q15

0x0000170A

Offset for finding a better pitch candidate in Q15 format

swCANDCORR_THLD1_Q15

0x799A

Pitch candidate correlation threshold 1 in Q15 format

swCANDCORR_THLD2_Q15

0x599A

Pitch candidate correlation threshold 2 in Q15 format

swCANDCORR_THLD3_Q15

0x6CCD

Pitch candidate correlation threshold 3 in Q15 format

swCANDAMP_THLD3_Q15

0x68F6

Pitch candidate amplitude threshold 3 in Q15 format

swSTARTFREQ_COEFF

0x553F

Start frequency coefficient (for candidate search)

swENDFREQ_COEFF

0x4666

End frequency coefficient (for candidate search)

DIRICHLET_KERNEL_SPAN

8

Direchlet kernal span (for interpolation)

REF_SAMPLE_RATE

8000

Reference sampling rate

REF_BANDWIDTH

4000

Reference bandwidth

lwTHIRD_REF_BANDWIDTH

87381333

One third of the reference bandwidth

lwTWO_THIRDS_REF_BANDWIDTH

174762667

Two thirds of the reference bandwidth

swCENTER_WEIGHT

0x5000

Center weight

swSIDE_WEIGHT

0x1800

Side weight

swAMP_SCALE_DOWN1

0x5333

Amplitude scale down factor 1

swAMP_SCALE_DOWN2

0x399A

Amplitude scale down factor 2

swAMP_SCALE_DOWN2b

0x7333

Amplitude scale down factor 2b

swUDIST1

-4160

Utility function distance 1

swUDIST2

-6400

Utility function distance 2

swUSTEP

-16384

Utility function step

swFREQ_MARGIN1

0x4AE1

Frequency margin 1

swAMP_MARGIN1

0x07AE

Amplitude margin 1

swAMP_MARGIN2

0x07AE

Amplitude margin 2

MIN_STABLE_FRAMES

6

Minimum number of stable frames

MAX_TRACK_GAP_FRAMES

2

Maximum pitch track gap frames

swSTABLE_FREQ_UPPER_MARGIN

0x4E14

Stable frequency upper margin

swSTABLE_FREQ_LOWER_MARGIN

0x68EB

Stable frequency lower margin

UNVOICED

0

Pitch frequency of an unvoiced frame

lwMAX_PITCH_FREQ

0x01A40000L

Maximum pitch frequency

lwMIN_PITCH_FREQ

0x00340000L

Minimum pitch frequency

MAX_PITCH_FREQ

420

Maximum pitch frequency in Hz

MIN_PITCH_FREQ

52

Minimum pitch frequency in Hz

HIGHPASS_CUTOFF_FREQ

300

Highpass cut-off frequency in Hz

NO_OF_FRACS

77

Number of fractions in the frations table

lwSHORT_WIN_START_FREQ

0x00C80000L

Short window start frequency

lwSHORT_WIN_END_FREQ

0x01A40000

Short window end frequency

lwSINGLE_WIN_START_FREQ

0x00640000L

Single window start frequency

lwSINGLE_WIN_END_FREQ

0x00D20000L

Single window end frequency

lwDOUBLE_WIN_START_FREQ

0x00340000

Double window start frequency

lwDOUBLE_WIN_END_FREQ

0x00780000L

Double window end frequency

MAX_LOCAL_MAXIMA_ON_SPECTRUM

70

Maximum number of local maxima on the spectrum

MAX_PEAKS_FOR_SORT

30

Maximum number peaks for sorting

MAX_PEAKS_PRELIM

7

Maximum number of peaks (preliminary)

MIN_PEAKS

7

Minimum number of peaks

MAX_PEAKS_FINAL

20

Maximum number of peaks (final)

MAX_PRELIM_CANDS

4

Maximum number of preliminary candidates (pitch)

CREATE_PIECEWISE_FUNC_LOOP_LIM_SH

20

Create Piecewise function loop limit for short window

CREATE_PIECEWISE_FUNC_LOOP_LIM_SNG

30

Create Piecewise function loop limit for single window

CREATE_PIECEWISE_FUNC_LOOP_LIM_DBL

60

Create Piecewise function loop limit for double window

swSUM_FRACTION

0x799A

Sum fraction

swAMP_FRACTION

0x33F8

Amplitude fraction

MAX_BEST_CANDS

2

Maximum number of best candidates (pitch)

N_OF_BEST_CANDS_SHORT

2

Number of best candidates for short window

N_OF_BEST_CANDS_SINGLE

2

Number of best candidates for single window

N_OF_BEST_CANDS_DOUBLE

2

Number of best candidates for double window

N_OF_BEST_CANDS

6

Number of best candidates for all windows

SIZE_SCRATCH_DOPITCH

1090

Scratch memory size for DoPitch() function (This is the actual size required. The declared size in C simulation is 1632)

SIZE_SCRATCH_ADVPROCESS

825

Scratch memory size for DoAdvProcess() function (This is the actual size required. The declared size in C simulation is 1100)

RVC_PITCH_ROM_SIG

11031

Signature for RVC_PITCH_ROM structure

RVC_PITCH_METER_SIG

21053

Signature for RVC_PITCH_METER structure

4.5.2 Description of fixed tables used in the C-code

This section contains a listing of all fixed tables sorted by source file name and table name. All table data is declared as Word16.

Table 6a: Fixed tables for AFE

File

Table Name

Length

Description

16kHzProcessing_B.c

table_pow2

33

Table for square root

LambdaNSEx2

100

Table used to compute first 100 LambdaNSE

dp02_h

59

MSB of QMF filter coefficients

dp02_l

43

LSB of QMF filter coefficients

PostProc_B.c

targetLMS16

12

Target for blind equalization

ComCeps_B.c

HalfHamming16

100

Hamming window coefficients

CosMatrix16

144

Inverse cosinus coefficients at 8Khz (not used at 16khz)

CosMatrix16_16khz

156

Inverse cosinus coefficients at 16Khz

pondMelFilter

309

Mel bank coefficients

ff4nrFix16_B.c

tabSin

64

Sine table

tabCos

64

Cosine table

MathFunc.c

tbInt0

48

Coefficients for computation of square root

ExtNoiseSup_B.c

lambda_1divX

20

Computation of 1/N

Hann_sh32_hi

100

MSB of hanning window coefficients (32 bits)

Hann_sh32_lo

100

LSB of hanning window coefficients (32 bits)

Hann_sh24_hi

100

MSB of hanning window coefficients (24 bits)

Hann_sh24_lo

100

LSB of hanning window coefficients (24 bits)

pondMelFilterNoise

157

Mel-frequency scale coefficients (applied to the Wiener filter)

idctMel16

234

Mel-warped inverse DCT coefficients

pondMelFilter16k

134

Filter bank coefficients at 16Khz

M1_LamdaLTE

8

Computation of 1/N

M1_LambdaNSEx2

100

Computation of 2/N

M1_LamdaNSE

9

Computation of 1/N

mInvLambda16

10

Comutation od 2/N

Table 6b: Fixed tables for VQ

File

Table Name

Length

Description

coder_VAD.c

quantizer16kHz_0_1

128

vq table

quantizer16kHz_2_3

128

vq table

quantizer16kHz_4_5

128

vq table

quantizer16kHz_6_7

128

vq table

quantizer16kHz_8_9

128

vq table

quantizer16kHz_10_11

64

vq table

quantizer16kHz_12_13

512

vq table

quantizer8kHz_0_1

128

vq table

quantizer8kHz_2_3

128

vq table

quantizer8kHz_4_5

128

vq table

quantizer8kHz_6_7

128

vq table

quantizer8kHz_8_9

128

vq table

quantizer8kHz_10_11

64

vq table

quantizer8kHz_12_13

512

vq table

weight16kHz_c0_shift

1

vq weights

weight16kHz_c0_norm

1

vq weights

weight16kHz_logE

1

vq weights

weight8kHz_c0_shift

1

vq weights

weight8kHz_c0_norm

1

vq weights

weight8kHz_logE

1

vq weights

plwQuantLevels[127]

127*2

vq tables for pitch/class quantization

ppplwQuantSections[8][3]

24*2

vq tables for pitch/class quantization

plwQuantLevels[31]

31*2

vq tables for pitch/class quantization

pplwQuantSections[4][3]

12*2

vq tables for pitch/class quantization

pswRatioThld_1[4][6]

24

vq tables for pitch/class quantization

piMultiLevelIndex[4]

4

vq tables for pitch/class quantization

pswRatioThld_2[4][8]

32

vq tables for pitch/class quantization

piMultiLevelIndex_2[4]

4

vq tables for pitch/class quantization

swAlpha1

1

pitch/class constants

swAlpha2

1

pitch/class constants

Table 6c: Fixed Tables for Extension

File

Table name

Length

Description

ExtNoiseSup_B.c

pswPePower

129

Coefficients to compute the pre-emphasis power spectrum

preProc_B.c

pswHpfCoef

15

High pass filter coefficients

preProc_B.c

pswLpfCoef

15

Low pass filter coefficients

preProc_B.c

pswLfeCoef

3

Low frequency emphasis filter coefficients

dsrAfeVad_B.c

piBurstConst

20

Burst length constants for different SNR’s

dsrAfeVad_B.c

piHangConst

20

Hang length constants for different SNR’s

dsrAfeVad_B.c

piVADThld

20

VAD voice metric thresholds for different SNR’s

dsrAfeVad_B.c

piVMTable

90

Voice metric table as a function of SNR index

dsrAfeVad_B.c

piSigThld

20

Signal threshold table as a function of SNR

dsrAfeVad_B.c

piUpdateThld

20

Update threshold table as a function of SNR

dsrAfeVad_B.c

pswShapeTable

23

Spectral shape correction table

fix_mathlib.c

coeff_sqrt5_58

5

Coefficients for computation of square root

fix_mathlib.c

coeff_sqrt5_78

5

Coefficients for computation of square root

rvc_pitch_init_B.h

ROM_astFrac

312

Fractions table

rvc_pitch_init_B.h

ROM_pstWindowshiftTable

514

Complex exponents table for time shifting in frequency domain

rvc_pitch_init_B.h

ROM_aswDirichletImag

8

Imaginary part of the Dirichlet kernel

4.5.3 Static variables used in the C-code

In this section two tables that specify the static variables for the AFE, VQ, and Extension respectively are shown.

Table 7a: AFE static variables

Struct Name

Variable

Type[Length]

Description

QMF_FIR

lengthQMF

Word32

QMF Filter length

*dp_l

Word16

QMF filter low frequency Coeff

*dp_h

Word16

QMF filter high frequency Coeff

*T

Word16

Temporary QMF filter buffer

T_dec

Word16

Multiplier for T

DataFor16kProc_B

FrameLength

Word32

Input Frame length

FrameShift

Word32

Shift value for the frame

numFramesInBuffer

Word32

Number of frames in buffer

SamplingFrequency

Word32

Sampling frequency (8/16)

Do16kHzProc

BOOLEAN

Flag to enable 16kHz processing

*hpBands_B

Word32

Buffer for HP bands

hpBandsSize

Word32

hpBands_B buffer size

CodeForBands16k_B

Word32[9]

HP coding buffer

bufferCodeForBands16k_B

Word32[27]

buffer used for HP coding

codeWeights_B

Word16[3]

code Weights buffer

bufferCodeWeights_B

Word16[9]

buffer used for code Weights

* pQMF_Fir

QMF_FIR

Pointer to QMF_FIR structure

*bufferData16k_B

Word32

temporary buffer to carry QMF LP data

bufData16kSize

Word32

16k data buffer size

*FirstWindow16k

MelFB_Window

pointer to MelFB_Window structure

noiseSE16k_B

Word32[3]

noise spectrul energy variable

noise_dec

Word16

Multiplier for noiseSE16k_B

BandsForCoding16k_B

Word32[9]

buffer for storing Bands for Coding

vadCounter16k

Word32

vad flag counter

vad16k

Word32

vad flag

nbSpeechFrames16k

Word32

number of speech frames counter

hangOver16k

Word32

hang over used for VAD

meanEn16k

Word32

mean Energy variable

nb_frame_threshold_nse

Word32

threshold NSE for frame

lambda_nse

Word16

lambda NSE variable

*dataHP_B

Word32

buffer stores QMF HP value

dec_16k

Word16[5]

Multiplier for dataHP_B buffer

BFC_dec

Word16[1]

Multiplier for computing bands for coding

fb16k_dec

Word16[3]

Buffer is used to store multiplier for current and pervious two frames

PostProcStructX

weightLMS

Word32[12]

Current LMS weight

CompCepsStructX

FFTLength

Word32

FFT size

Do16khzProc

Word16

Flag to enable 16kHz processing

*pData16k

Word32

Pointer to data for 16Khz processing

WaveProcStructX

*TeagerFilter16

Word32

Pointer to teager filter

*TeagerWindow32

Word32

Pointer to teager window

TeagerOnset

Word32

Unused

FrameLength

Word32

Input frame length

ns_var_F

SampFreq

Word16

Sampling frequency (8/16)

Do16khzProc

Word16

Flag to enable 16kHz processing

buffers.nbFramesInFirstStage

Word32

number of frames in first stage

buffers.nbFramesInFirstStage

Word32

number of frames in second stage

buffers. nbFramesOutSecondStage

Word32

number of frames out og second stage

buffers. FirstStageIn16Buffer

Word16[180]

First stage buffer

buffers.SecondStageInBuffer32

Word32[180]

Second stage buffer

buffers. SecondDecalSig

Word16[4]

Shift factor for each sub-frame of second stage buffer

prevSamples32.lastSampleIn32

Word32

Last input sample of DC offset compensation

prevSamples32.lastDCOut32

Word32

last output sample of DC offset compensation

prevSamples32. oldShift

Word16

lprevious window shift factor of DC offset compensation

spectrum.indexBuffer1

Word16

Where to enter new PSD for first stage, alternatively 0 and 1

spectrum.indexBuffer2

Word16

Where to enter new PSD for second stage, alternatively 0 and 1

spectrum.noiseSE1_32

Word32[65]

Noise spectrum estimate for first stage

spectrum.noiseSE1_dec

Word16[65]

Shift factor for Noise spectrum estimate (first sage)

spectrum.noiseSE2_32

Word32[65]

Noise spectrum estimate for second stage

spectrum.noiseSE2_dec

Word16[65]

Shift factor for Noise spectrum estimate (second sage)

spectrum.PSDMeanAntBuffer1

Word32[65]

1st stage PSD Mean buffer for precedent frame

spectrum.nSigSE1Ant_dec

Word16[65]

Shift factor for PSD Mean buffer for precedent frame (1rst stage)

spectrum.PSDMeanAntBuffer2

Word32[65]

2nd stage PSD Mean bufferfor precedent frame

spectrum.nSigSE2Ant_dec

Word16[65]

Shift factor for PSD Mean buffer for precedent frame (2nd stage)

spectrum.denSigSE1_32

Word32[65]

1st stage PSD Mean buffer

spectrum. nSigSE1Cur_dec

Word16[65]

Shift factor for PSD Mean buffer (1rst stage)

spectrum. denSigSE2_32

Word32[65]

2nd stage PSD Mean buffer

spectrum. nSigSE2Cur_dec

Word16[65]

Shift factor for PSD Mean buffer (2nd stage)

vad_data_ns_F. nbFrame

Word16[2]

Nubmer of frames (for the 2 stages)

vad_data_ns_F. flagVAD

Word16

Vad Flag (1 = SPEECH, 0 = NON SPEECH)

vad_data_ns_F.hangOver

Word16

hangover

vad_data_ns_F. nbSpeechFrames

Word16

Number of speech frames (used to set hangover)

vad_data_ns_F.meanEn32

Word32

Mean energy for VAD

vad_data_ca. flagVAD

Word16

Vad Flag (1 = SPEECH, 0 = NON SPEECH)

vad_data_ca.hangOver

Word16

hangover

vad_data_ca. nbSpeechFrames

Word16

Number of speech frames (used to set hangover)

vad_data_ca.meanEn32

Word32

Mean energy for VAD

vad_data_fd.MelMean

Word16

SpeechQMel (for frame dropping)

vad_data_fd.VarMean

Word32

SpeechQVar (for frame dropping)

vad_data_fd.AccTest

Word32

SpeechQSpec (for frame dropping)

vad_data_fd.AccTest2

Word32

vad_data_fd.SpecMean

Word32

SpecMean (for frame dropping)

vad_data_fd.MelValues

Word16[2]

SpeechQMel (for frame dropping)

vad_data_fd.SpecValues

Word32

SpeechQSpec (for frame dropping)

vad_data_fd.SpeechInVADQ

Word16

Flag (for frame dropping)

vad_data_fd.SpeechInVADQ2

Word16

Flag (for frame dropping)

gainFact.logDenEn1_32

Word32[3]

Denoise frame energy for gain factorization

gainFact.lowSNRtrack32

Word32

Low SNR level for gain factorization

gainFact. alfaGF16

Word16

Wiener filter gain factorization coefficient

VADStructX_F

Focus

Word16

Position of circular buffe

HangOver

Word16

Hangover length

FlushFocus

Word16

Position in circular buffer when emptying at end

H_CountDown

Word16

Main hangover countdown

V_CountDown

Word16

Short hangover countdown

**OutBuffer

Word32

outBuffer pointer pointer

*OutBuffer

Word32[7]

outBuffer pointer

OutBuffer

Word16[7×15]

outBuffer

Table 7b: VQ static variables

Struct Name

Variable

Type [Length]

Description

coder_VAD.c

four_frames[27]

Word16[27]

Previous frames used to build multiframe

plwQPHistory[3]

Word32[3]

History of Pitch

IReliableFlag

Word16

Pitch reliability flag

Table 7c: Extension static variables

Struct Name

Variable

Type[Length]

Description

iFirstFrameFlag

Word16

First frame flag

pswUBSpeech

Word16[200]

Upper band speech

pswDownSampledProcSpeech

Word16[75]

Down-sampled processed speech

lwCritMax

Word32

Maximum power ratio

iOldPitchPeriod

Word16

Old pitch period value

iOldFrameNo

Word16

Old frame number

PCORR_STATE_be

s_be

lwX1_X1

Word32

X1*X1

lwZ1_Z1

Word32

Z1*Z1

lwZ2_Z2

Word32

Z2*Z2

lwX1_Z1

Word32

X1*Z1

lwX1_Z2

Word32

X1*Z2

lwZ1_Z2

Word32

Z1*Z2

swX1_Sum

Word16

Sum of X1

swZ1_Sum

Word16

Sum of Z1

swZ2_Sum

Word16

Sum of Z2

iBurstConst

Word16

Burst constant

iBurstCount

Word16

Burst count

iHangConst

Word16

Hang constant

iHangCount

Word16

Hang count

iVADThld

Word16

VAD threshold

iFrameCount

Word16

Frame count

iFUpdateFlag

Word16

Forced update flag

iHysterCount

Word16

Hysteresis count

iLastUpdateCount

Word16

Last update count

iSigThld

Word16

Signal threshold

iUpdateCount

Word16

Update count

iChanEnrgShift

Word16

Channel energy shift

iChanNoiseEnrgShift

Word16

Channel noise energy shift

pswChanEnrg

Word16[23]

Channel energy

pswChanNoiseEnrg

Word16[23]

Channel noise energy

swBeta

Word16

Beta value

swSnr

Word16

SNR value

NormSw

pnsLogSpecEnrgLong

swMantissa

Word16[23]

Mantissa

iShift

Word16[23]

Shift

swC0

Word16

C0 value

swC1

Word16

C1 value

swC2

Word16

C2 value

pswHpfXState

Word16[6]

High pass filter input state

pswHpfYState

Word16[12]

High pass filter output state

pswLpfXState

Word16[6]

Low pass filter input state

pswLpfYState

Word16[12]

Low pass filter output state

pswLfeXState

Word16

Low frequency emphasis filter input state

pswLfeYState

Word16[2]

Low frequency emphasis filter output state