vad2.c

arm音频编解码库

/*
*****************************************************************************
*                                                                        
*      GSM AMR-NB speech codec   R98   Version 7.6.0   December 12, 2001
*                                R99   Version 3.3.0                
*                                REL-4 Version 4.1.0                
*
*****************************************************************************
*
*      File             : vad2.c
*      Purpose          : Voice Activity Detection (VAD) for AMR (option 2)
*
*****************************************************************************
*/

/*
*****************************************************************************
*                         VERSION ID
*****************************************************************************
*/

const char vad2_id[] = "@(#)$Id $";


/***************************************************************************
 *
 *   FUNCTION NAME: vad2()
 *
 *   PURPOSE:
 *     This function provides the Voice Activity Detection function option 2
 *     for the Adaptive Multi-rate (AMR) codec.
 *
 *   INPUTS:
 *
 *     farray_ptr
 *                     pointer to Word16[80] input array
 *     vadState2
 *                     pointer to vadState2 state structure
 *
 *   OUTPUTS:
 *
 *     state variables are updated
 *
 *   RETURN VALUE:
 *
 *     Word16
 *                     VAD(m) - two successive calls to vad2() yield
 *                     the VAD decision for the 20 ms frame:
 *                     VAD_flag = VAD(m-1) || VAD(m)
 *
 *
 *************************************************************************/

/* Includes */

#include <stdio.h>
#include <stdlib.h>

#include "typedef.h"
#include "cnst.h"
#include "basic_op.h"
#include "oper_32b.h"
#include "count.h"
#include "log2.h"
#include "pow2.h"

#include "vad2.h"


/* Local functions */

/***************************************************************************
 *
 *   FUNCTION NAME: fn10Log10
 *
 *   PURPOSE:
 *     The purpose of this function is to take the 10*log base 10 of input and
 *     divide by 128 and return; i.e. output = 10*log10(input)/128 (scaled as 7,8)
 *
 *   INPUTS:
 *
 *     L_Input
 *                     input (scaled as 31-fbits,fbits)
 *     fbits
 *                     number of fractional bits on input
 *
 *   OUTPUTS:
 *
 *     none
 *
 *   RETURN VALUE:
 *
 *     Word16
 *                     output (scaled as 7,8)
 *
 *   DESCRIPTION:
 *
 *     10*log10(x)/128 = 10*(log10(2) * (log2(x<<fbits)-log2(1<<fbits)) >> 7
 *                     = 3.0103 * (log2(x<<fbits) - fbits) >> 7
 *                     = ((3.0103/4.0 * (log2(x<<fbits) - fbits) << 2) >> 7
 *                     = (3.0103/4.0 * (log2(x<<fbits) - fbits) >> 5
 *
 *************************************************************************/

Word16 fn10Log10 (Word32 L_Input, Word16 fbits)
{

	Word16 integer;		/* Integer part of Log2.   (range: 0<=val<=30) */
	Word16 fraction;	/* Fractional part of Log2. (range: 0<=val<1) */

	Word32 Ltmp;
	Word16 tmp;

        Log2(L_Input, &integer, &fraction);

	integer = sub(integer, fbits);
	Ltmp = Mpy_32_16 (integer, fraction, 24660);	/* 24660 = 10*log10(2)/4 scaled 0,15 */
	Ltmp = L_shr_r(Ltmp, 5+1);			/* extra shift for 30,1 => 15,0 extract correction */
        tmp = extract_l(Ltmp);

        return (tmp);
}


/***************************************************************************
 *
 *   FUNCTION NAME: block_norm
 *
 *   PURPOSE:
 *     The purpose of this function is block normalise the input data sequence
 *
 *   INPUTS:
 *
 *     &in[0]
 *                     pointer to data sequence to be normalised
 *     length
 *                     number of elements in data sequence
 *     headroom
 *                     number of headroom bits (i.e., 
 *
 *   OUTPUTS:
 *
 *     &out[0]
 *                     normalised output data sequence pointed to by &out[0]
 *
 *   RETURN VALUE:
 *
 *     Word16
 *                     number of bits sequence was left shifted
 *
 *   DESCRIPTION:
 *
 *                     1) Search for maximum absolute valued data element
 *                     2) Normalise the max element with "headroom"
 *                     3) Transfer/shift the input sequence to the output buffer
 *                     4) Return the number of left shifts
 *
 *   CAVEATS:
 *                     An input sequence of all zeros will return the maximum
 *                     number of left shifts allowed, NOT the value returned
 *                     by a norm_s(0) call, since it desired to associate an
 *                     all zeros sequence with low energy.
 *
 *************************************************************************/

Word16 block_norm (Word16 * in, Word16 * out, Word16 length, Word16 headroom)
{

	Word16 i, max, scnt, adata;

        max = abs_s(in[0]);
	for (i = 1; i < length; i++)
	{
                adata = abs_s(in[i]);                           test();
		if (sub(adata, max) > 0)
		{
			max = adata;				move16();
		}
	}
	test();
	if (max != 0)
	{
		scnt = sub(norm_s(max), headroom);
		for (i = 0; i < length; i++)
		{
			out[i] = shl(in[i], scnt);	       	move16();
		}
	}
	else
	{
		scnt = sub(16, headroom);
		for (i = 0; i < length; i++)
		{
			out[i] = 0;                             move16();
		}
	}
	return (scnt);
}



/********************************************* The VAD function ***************************************************/

Word16 vad2 (Word16 * farray_ptr, vadState2 * st)
{

	/*
	 * The channel table is defined below.  In this table, the
	 * lower and higher frequency coefficients for each of the 16
	 * channels are specified.  The table excludes the coefficients
	 * with numbers 0 (DC), 1, and 64 (Foldover frequency).
	 */

	static Word16 ch_tbl[NUM_CHAN][2] =
	{

		{2, 3},
		{4, 5},
		{6, 7},
		{8, 9},
		{10, 11},
		{12, 13},
		{14, 16},
		{17, 19},
		{20, 22},
		{23, 26},
		{27, 30},
		{31, 35},
		{36, 41},
		{42, 48},
		{49, 55},
		{56, 63}

	};

	/* channel energy scaling table - allows efficient division by number
         * of DFT bins in the channel: 1/2, 1/3, 1/4, etc.
	 */

	static Word16 ch_tbl_sh[NUM_CHAN] =
	{
		16384, 16384, 16384, 16384, 16384, 16384, 10923, 10923,
		10923, 8192, 8192, 6554, 5461, 4681, 4681, 4096
	};

	/*
	 * The voice metric table is defined below.  It is a non-
	 * linear table with a deadband near zero.  It maps the SNR
	 * index (quantized SNR value) to a number that is a measure
	 * of voice quality.
	 */

	static Word16 vm_tbl[90] =
	{
		2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
		3, 3, 3, 3, 3, 4, 4, 4, 5, 5, 5, 6, 6, 7, 7, 7,
		8, 8, 9, 9, 10, 10, 11, 12, 12, 13, 13, 14, 15,
		15, 16, 17, 17, 18, 19, 20, 20, 21, 22, 23, 24,
		24, 25, 26, 27, 28, 28, 29, 30, 31, 32, 33, 34,
		35, 36, 37, 37, 38, 39, 40, 41, 42, 43, 44, 45,
		46, 47, 48, 49, 50, 50, 50, 50, 50, 50, 50, 50,
		50, 50
	};

	/* hangover as a function of peak SNR (3 dB steps) */
	static Word16 hangover_table[20] =
	{
		30, 30, 30, 30, 30, 30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 8, 8, 8
	};

	/* burst sensitivity as a function of peak SNR (3 dB steps) */
	static Word16 burstcount_table[20] =
	{
		8, 8, 8, 8, 8, 8, 8, 8, 7, 6, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4
	};

	/* voice metric sensitivity as a function of peak SNR (3 dB steps) */
	static Word16 vm_threshold_table[20] =
	{
                34, 34, 34, 34, 34, 34, 34, 34, 34, 34, 34, 40, 51, 71, 100, 139, 191, 257, 337, 432
	};


	/* State tables that use 22,9 or 27,4 scaling for ch_enrg[] */

	static Word16 noise_floor_chan[2] =	{NOISE_FLOOR_CHAN_0, NOISE_FLOOR_CHAN_1};
	static Word16 min_chan_enrg[2] =	{MIN_CHAN_ENRG_0, MIN_CHAN_ENRG_1};
	static Word16 ine_noise[2] = 		{INE_NOISE_0, INE_NOISE_1};
	static Word16 fbits[2] = 		{FRACTIONAL_BITS_0, FRACTIONAL_BITS_1};
	static Word16 state_change_shift_r[2] = {STATE_1_TO_0_SHIFT_R, STATE_0_TO_1_SHIFT_R};

	/* Energy scale table given 30,1 input scaling (also account for -6 dB shift on input) */
	static Word16 enrg_norm_shift[2] = 	{(FRACTIONAL_BITS_0-1+2), (FRACTIONAL_BITS_1-1+2)};


	/* Automatic variables */

	Word32 Lenrg;				/* scaled as 30,1 */
	Word32 Ltne;				/* scaled as 22,9 */
	Word32 Ltce;				/* scaled as 22,9 or 27,4 */

	Word16 tne_db;				/* scaled as 7,8 */
	Word16 tce_db;				/* scaled as 7,8 */

	Word16 input_buffer[FRM_LEN];		/* used for block normalising input data */
	Word16 data_buffer[FFT_LEN];		/* used for in-place FFT */

	Word16 ch_snr[NUM_CHAN];		/* scaled as 7,8 */
	Word16 ch_snrq;				/* scaled as 15,0 (in 0.375 dB steps) */
	Word16 vm_sum;				/* scaled as 15,0 */
	Word16 ch_enrg_dev;			/* scaled as 7,8 */

	Word32 Lpeak;				/* maximum channel energy */
	Word16 p2a_flag;			/* flag to indicate spectral peak-to-average ratio > 10 dB */

	Word16 ch_enrg_db[NUM_CHAN];		/* scaled as 7,8 */
	Word16 ch_noise_db;			/* scaled as 7,8 */

	Word16 alpha;				/* scaled as 0,15 */
	Word16 one_m_alpha;			/* scaled as 0,15 */
	Word16 update_flag;			/* set to indicate a background noise estimate update */

	Word16 i, j, j1, j2;			/* Scratch variables */
	Word16 hi1, lo1;

	Word32 Ltmp, Ltmp1, Ltmp2;
	Word16 tmp;

	Word16 normb_shift;		/* block norm shift count */

	Word16 ivad;			/* intermediate VAD decision (return value) */
	Word16 tsnrq;			/* total signal-to-noise ratio (quantized 3 dB steps) scaled as 15,0 */
	Word16 xt;			/* instantaneous frame SNR in dB, scaled as 7,8 */

	Word16 state_change;


	/* Increment frame counter */
	st->Lframe_cnt = L_add(st->Lframe_cnt, 1);

	/* Block normalize the input */
	normb_shift = block_norm(farray_ptr, input_buffer, FRM_LEN, FFT_HEADROOM);

	/* Pre-emphasize the input data and store in the data buffer with the appropriate offset */
	for (i = 0; i < DELAY; i++)
	{
		data_buffer[i] = 0;									move16();
	}

	st->pre_emp_mem = shr_r(st->pre_emp_mem, sub(st->last_normb_shift, normb_shift));
	st->last_normb_shift = normb_shift;								move16();

	data_buffer[DELAY] = add(input_buffer[0], mult(PRE_EMP_FAC, st->pre_emp_mem));			move16();

	for (i = DELAY + 1, j = 1; i < DELAY + FRM_LEN; i++, j++)
	{
		data_buffer[i] = add(input_buffer[j], mult(PRE_EMP_FAC, input_buffer[j-1]));		move16();
	}
	st->pre_emp_mem = input_buffer[FRM_LEN-1];							move16();

	for (i = DELAY + FRM_LEN; i < FFT_LEN; i++)
	{
		data_buffer[i] = 0;									move16();
	}


	/* Perform FFT on the data buffer */
	r_fft(data_buffer);


	/* Use normb_shift factor to determine the scaling of the energy estimates */
	state_change = 0;										move16();
													test();
	if (st->shift_state == 0)
	{												test();
		if (sub(normb_shift, -FFT_HEADROOM+2) <= 0)
		{
			state_change = 1;								move16();
			st->shift_state = 1;								move16();
		}
	}
	else
	{												test();
		if (sub(normb_shift, -FFT_HEADROOM+5) >= 0)
		{
			state_change = 1;								move16();
			st->shift_state = 0;								move16();
		}
	}

	/* Scale channel energy estimate */								test();
	if (state_change)
	{
		for (i = LO_CHAN; i <= HI_CHAN; i++)
		{
			st->Lch_enrg[i] = L_shr(st->Lch_enrg[i], state_change_shift_r[st->shift_state]);	move32();
		}
	}


	/* Estimate the energy in each channel */
													test();
	if (L_sub(st->Lframe_cnt, 1) == 0)
	{
		alpha = 32767;										move16();
		one_m_alpha = 0;									move16();
	}
	else
	{
		alpha = CEE_SM_FAC;									move16();
		one_m_alpha = ONE_MINUS_CEE_SM_FAC;							move16();
	}

	for (i = LO_CHAN; i <= HI_CHAN; i++)
	{
		Lenrg = 0;										move16();
		j1 = ch_tbl[i][0];									move16();
		j2 = ch_tbl[i][1];									move16();

		for (j = j1; j <= j2; j++)
		{
			Lenrg = L_mac(Lenrg, data_buffer[2 * j], data_buffer[2 * j]);
			Lenrg = L_mac(Lenrg, data_buffer[2 * j + 1], data_buffer[2 * j + 1]);
		}

		/* Denorm energy & scale 30,1 according to the state */
		Lenrg = L_shr_r(Lenrg, sub(shl(normb_shift, 1), enrg_norm_shift[st->shift_state]));

		/* integrate over time: e[i] = (1-alpha)*e[i] + alpha*enrg/num_bins_in_chan */
		tmp = mult(alpha, ch_tbl_sh[i]);
		L_Extract (Lenrg, &hi1, &lo1);
		Ltmp = Mpy_32_16(hi1, lo1, tmp);

		L_Extract (st->Lch_enrg[i], &hi1, &lo1);
		st->Lch_enrg[i] = L_add(Ltmp, Mpy_32_16(hi1, lo1, one_m_alpha));			move32();
													test();
		if (L_sub(st->Lch_enrg[i], min_chan_enrg[st->shift_state]) < 0)
		{
			st->Lch_enrg[i] = min_chan_enrg[st->shift_state];				move32();
		}

	}


	/* Compute the total channel energy estimate (Ltce) */
	Ltce = 0;											move16();
	for (i = LO_CHAN; i <= HI_CHAN; i++)
	{
		Ltce = L_add(Ltce, st->Lch_enrg[i]);
	}


	/* Calculate spectral peak-to-average ratio, set flag if p2a > 10 dB */
	Lpeak = 0;											move32();
	for (i = LO_CHAN+2; i <= HI_CHAN; i++)	/* Sine waves not valid for low frequencies */
	{												test();
		if (L_sub(st->Lch_enrg [i], Lpeak) > 0)
		{