cannam@154: /*********************************************************************** cannam@154: Copyright (c) 2006-2011, Skype Limited. All rights reserved. cannam@154: Redistribution and use in source and binary forms, with or without cannam@154: modification, are permitted provided that the following conditions cannam@154: are met: cannam@154: - Redistributions of source code must retain the above copyright notice, cannam@154: this list of conditions and the following disclaimer. cannam@154: - Redistributions in binary form must reproduce the above copyright cannam@154: notice, this list of conditions and the following disclaimer in the cannam@154: documentation and/or other materials provided with the distribution. cannam@154: - Neither the name of Internet Society, IETF or IETF Trust, nor the cannam@154: names of specific contributors, may be used to endorse or promote cannam@154: products derived from this software without specific prior written cannam@154: permission. cannam@154: THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" cannam@154: AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE cannam@154: IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE cannam@154: ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE cannam@154: LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR cannam@154: CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF cannam@154: SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS cannam@154: INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN cannam@154: CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) cannam@154: ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE cannam@154: POSSIBILITY OF SUCH DAMAGE. cannam@154: ***********************************************************************/ cannam@154: cannam@154: #ifdef HAVE_CONFIG_H cannam@154: #include "config.h" cannam@154: #endif cannam@154: cannam@154: #include "main.h" cannam@154: #include "stack_alloc.h" cannam@154: cannam@154: /* Silk VAD noise level estimation */ cannam@154: # if !defined(OPUS_X86_MAY_HAVE_SSE4_1) cannam@154: static OPUS_INLINE void silk_VAD_GetNoiseLevels( cannam@154: const opus_int32 pX[ VAD_N_BANDS ], /* I subband energies */ cannam@154: silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */ cannam@154: ); cannam@154: #endif cannam@154: cannam@154: /**********************************/ cannam@154: /* Initialization of the Silk VAD */ cannam@154: /**********************************/ cannam@154: opus_int silk_VAD_Init( /* O Return value, 0 if success */ cannam@154: silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */ cannam@154: ) cannam@154: { cannam@154: opus_int b, ret = 0; cannam@154: cannam@154: /* reset state memory */ cannam@154: silk_memset( psSilk_VAD, 0, sizeof( silk_VAD_state ) ); cannam@154: cannam@154: /* init noise levels */ cannam@154: /* Initialize array with approx pink noise levels (psd proportional to inverse of frequency) */ cannam@154: for( b = 0; b < VAD_N_BANDS; b++ ) { cannam@154: psSilk_VAD->NoiseLevelBias[ b ] = silk_max_32( silk_DIV32_16( VAD_NOISE_LEVELS_BIAS, b + 1 ), 1 ); cannam@154: } cannam@154: cannam@154: /* Initialize state */ cannam@154: for( b = 0; b < VAD_N_BANDS; b++ ) { cannam@154: psSilk_VAD->NL[ b ] = silk_MUL( 100, psSilk_VAD->NoiseLevelBias[ b ] ); cannam@154: psSilk_VAD->inv_NL[ b ] = silk_DIV32( silk_int32_MAX, psSilk_VAD->NL[ b ] ); cannam@154: } cannam@154: psSilk_VAD->counter = 15; cannam@154: cannam@154: /* init smoothed energy-to-noise ratio*/ cannam@154: for( b = 0; b < VAD_N_BANDS; b++ ) { cannam@154: psSilk_VAD->NrgRatioSmth_Q8[ b ] = 100 * 256; /* 100 * 256 --> 20 dB SNR */ cannam@154: } cannam@154: cannam@154: return( ret ); cannam@154: } cannam@154: cannam@154: /* Weighting factors for tilt measure */ cannam@154: static const opus_int32 tiltWeights[ VAD_N_BANDS ] = { 30000, 6000, -12000, -12000 }; cannam@154: cannam@154: /***************************************/ cannam@154: /* Get the speech activity level in Q8 */ cannam@154: /***************************************/ cannam@154: opus_int silk_VAD_GetSA_Q8_c( /* O Return value, 0 if success */ cannam@154: silk_encoder_state *psEncC, /* I/O Encoder state */ cannam@154: const opus_int16 pIn[] /* I PCM input */ cannam@154: ) cannam@154: { cannam@154: opus_int SA_Q15, pSNR_dB_Q7, input_tilt; cannam@154: opus_int decimated_framelength1, decimated_framelength2; cannam@154: opus_int decimated_framelength; cannam@154: opus_int dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s; cannam@154: opus_int32 sumSquared, smooth_coef_Q16; cannam@154: opus_int16 HPstateTmp; cannam@154: VARDECL( opus_int16, X ); cannam@154: opus_int32 Xnrg[ VAD_N_BANDS ]; cannam@154: opus_int32 NrgToNoiseRatio_Q8[ VAD_N_BANDS ]; cannam@154: opus_int32 speech_nrg, x_tmp; cannam@154: opus_int X_offset[ VAD_N_BANDS ]; cannam@154: opus_int ret = 0; cannam@154: silk_VAD_state *psSilk_VAD = &psEncC->sVAD; cannam@154: SAVE_STACK; cannam@154: cannam@154: /* Safety checks */ cannam@154: silk_assert( VAD_N_BANDS == 4 ); cannam@154: celt_assert( MAX_FRAME_LENGTH >= psEncC->frame_length ); cannam@154: celt_assert( psEncC->frame_length <= 512 ); cannam@154: celt_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) ); cannam@154: cannam@154: /***********************/ cannam@154: /* Filter and Decimate */ cannam@154: /***********************/ cannam@154: decimated_framelength1 = silk_RSHIFT( psEncC->frame_length, 1 ); cannam@154: decimated_framelength2 = silk_RSHIFT( psEncC->frame_length, 2 ); cannam@154: decimated_framelength = silk_RSHIFT( psEncC->frame_length, 3 ); cannam@154: /* Decimate into 4 bands: cannam@154: 0 L 3L L 3L 5L cannam@154: - -- - -- -- cannam@154: 8 8 2 4 4 cannam@154: cannam@154: [0-1 kHz| temp. |1-2 kHz| 2-4 kHz | 4-8 kHz | cannam@154: cannam@154: They're arranged to allow the minimal ( frame_length / 4 ) extra cannam@154: scratch space during the downsampling process */ cannam@154: X_offset[ 0 ] = 0; cannam@154: X_offset[ 1 ] = decimated_framelength + decimated_framelength2; cannam@154: X_offset[ 2 ] = X_offset[ 1 ] + decimated_framelength; cannam@154: X_offset[ 3 ] = X_offset[ 2 ] + decimated_framelength2; cannam@154: ALLOC( X, X_offset[ 3 ] + decimated_framelength1, opus_int16 ); cannam@154: cannam@154: /* 0-8 kHz to 0-4 kHz and 4-8 kHz */ cannam@154: silk_ana_filt_bank_1( pIn, &psSilk_VAD->AnaState[ 0 ], cannam@154: X, &X[ X_offset[ 3 ] ], psEncC->frame_length ); cannam@154: cannam@154: /* 0-4 kHz to 0-2 kHz and 2-4 kHz */ cannam@154: silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState1[ 0 ], cannam@154: X, &X[ X_offset[ 2 ] ], decimated_framelength1 ); cannam@154: cannam@154: /* 0-2 kHz to 0-1 kHz and 1-2 kHz */ cannam@154: silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState2[ 0 ], cannam@154: X, &X[ X_offset[ 1 ] ], decimated_framelength2 ); cannam@154: cannam@154: /*********************************************/ cannam@154: /* HP filter on lowest band (differentiator) */ cannam@154: /*********************************************/ cannam@154: X[ decimated_framelength - 1 ] = silk_RSHIFT( X[ decimated_framelength - 1 ], 1 ); cannam@154: HPstateTmp = X[ decimated_framelength - 1 ]; cannam@154: for( i = decimated_framelength - 1; i > 0; i-- ) { cannam@154: X[ i - 1 ] = silk_RSHIFT( X[ i - 1 ], 1 ); cannam@154: X[ i ] -= X[ i - 1 ]; cannam@154: } cannam@154: X[ 0 ] -= psSilk_VAD->HPstate; cannam@154: psSilk_VAD->HPstate = HPstateTmp; cannam@154: cannam@154: /*************************************/ cannam@154: /* Calculate the energy in each band */ cannam@154: /*************************************/ cannam@154: for( b = 0; b < VAD_N_BANDS; b++ ) { cannam@154: /* Find the decimated framelength in the non-uniformly divided bands */ cannam@154: decimated_framelength = silk_RSHIFT( psEncC->frame_length, silk_min_int( VAD_N_BANDS - b, VAD_N_BANDS - 1 ) ); cannam@154: cannam@154: /* Split length into subframe lengths */ cannam@154: dec_subframe_length = silk_RSHIFT( decimated_framelength, VAD_INTERNAL_SUBFRAMES_LOG2 ); cannam@154: dec_subframe_offset = 0; cannam@154: cannam@154: /* Compute energy per sub-frame */ cannam@154: /* initialize with summed energy of last subframe */ cannam@154: Xnrg[ b ] = psSilk_VAD->XnrgSubfr[ b ]; cannam@154: for( s = 0; s < VAD_INTERNAL_SUBFRAMES; s++ ) { cannam@154: sumSquared = 0; cannam@154: for( i = 0; i < dec_subframe_length; i++ ) { cannam@154: /* The energy will be less than dec_subframe_length * ( silk_int16_MIN / 8 ) ^ 2. */ cannam@154: /* Therefore we can accumulate with no risk of overflow (unless dec_subframe_length > 128) */ cannam@154: x_tmp = silk_RSHIFT( cannam@154: X[ X_offset[ b ] + i + dec_subframe_offset ], 3 ); cannam@154: sumSquared = silk_SMLABB( sumSquared, x_tmp, x_tmp ); cannam@154: cannam@154: /* Safety check */ cannam@154: silk_assert( sumSquared >= 0 ); cannam@154: } cannam@154: cannam@154: /* Add/saturate summed energy of current subframe */ cannam@154: if( s < VAD_INTERNAL_SUBFRAMES - 1 ) { cannam@154: Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], sumSquared ); cannam@154: } else { cannam@154: /* Look-ahead subframe */ cannam@154: Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], silk_RSHIFT( sumSquared, 1 ) ); cannam@154: } cannam@154: cannam@154: dec_subframe_offset += dec_subframe_length; cannam@154: } cannam@154: psSilk_VAD->XnrgSubfr[ b ] = sumSquared; cannam@154: } cannam@154: cannam@154: /********************/ cannam@154: /* Noise estimation */ cannam@154: /********************/ cannam@154: silk_VAD_GetNoiseLevels( &Xnrg[ 0 ], psSilk_VAD ); cannam@154: cannam@154: /***********************************************/ cannam@154: /* Signal-plus-noise to noise ratio estimation */ cannam@154: /***********************************************/ cannam@154: sumSquared = 0; cannam@154: input_tilt = 0; cannam@154: for( b = 0; b < VAD_N_BANDS; b++ ) { cannam@154: speech_nrg = Xnrg[ b ] - psSilk_VAD->NL[ b ]; cannam@154: if( speech_nrg > 0 ) { cannam@154: /* Divide, with sufficient resolution */ cannam@154: if( ( Xnrg[ b ] & 0xFF800000 ) == 0 ) { cannam@154: NrgToNoiseRatio_Q8[ b ] = silk_DIV32( silk_LSHIFT( Xnrg[ b ], 8 ), psSilk_VAD->NL[ b ] + 1 ); cannam@154: } else { cannam@154: NrgToNoiseRatio_Q8[ b ] = silk_DIV32( Xnrg[ b ], silk_RSHIFT( psSilk_VAD->NL[ b ], 8 ) + 1 ); cannam@154: } cannam@154: cannam@154: /* Convert to log domain */ cannam@154: SNR_Q7 = silk_lin2log( NrgToNoiseRatio_Q8[ b ] ) - 8 * 128; cannam@154: cannam@154: /* Sum-of-squares */ cannam@154: sumSquared = silk_SMLABB( sumSquared, SNR_Q7, SNR_Q7 ); /* Q14 */ cannam@154: cannam@154: /* Tilt measure */ cannam@154: if( speech_nrg < ( (opus_int32)1 << 20 ) ) { cannam@154: /* Scale down SNR value for small subband speech energies */ cannam@154: SNR_Q7 = silk_SMULWB( silk_LSHIFT( silk_SQRT_APPROX( speech_nrg ), 6 ), SNR_Q7 ); cannam@154: } cannam@154: input_tilt = silk_SMLAWB( input_tilt, tiltWeights[ b ], SNR_Q7 ); cannam@154: } else { cannam@154: NrgToNoiseRatio_Q8[ b ] = 256; cannam@154: } cannam@154: } cannam@154: cannam@154: /* Mean-of-squares */ cannam@154: sumSquared = silk_DIV32_16( sumSquared, VAD_N_BANDS ); /* Q14 */ cannam@154: cannam@154: /* Root-mean-square approximation, scale to dBs, and write to output pointer */ cannam@154: pSNR_dB_Q7 = (opus_int16)( 3 * silk_SQRT_APPROX( sumSquared ) ); /* Q7 */ cannam@154: cannam@154: /*********************************/ cannam@154: /* Speech Probability Estimation */ cannam@154: /*********************************/ cannam@154: SA_Q15 = silk_sigm_Q15( silk_SMULWB( VAD_SNR_FACTOR_Q16, pSNR_dB_Q7 ) - VAD_NEGATIVE_OFFSET_Q5 ); cannam@154: cannam@154: /**************************/ cannam@154: /* Frequency Tilt Measure */ cannam@154: /**************************/ cannam@154: psEncC->input_tilt_Q15 = silk_LSHIFT( silk_sigm_Q15( input_tilt ) - 16384, 1 ); cannam@154: cannam@154: /**************************************************/ cannam@154: /* Scale the sigmoid output based on power levels */ cannam@154: /**************************************************/ cannam@154: speech_nrg = 0; cannam@154: for( b = 0; b < VAD_N_BANDS; b++ ) { cannam@154: /* Accumulate signal-without-noise energies, higher frequency bands have more weight */ cannam@154: speech_nrg += ( b + 1 ) * silk_RSHIFT( Xnrg[ b ] - psSilk_VAD->NL[ b ], 4 ); cannam@154: } cannam@154: cannam@154: if( psEncC->frame_length == 20 * psEncC->fs_kHz ) { cannam@154: speech_nrg = silk_RSHIFT32( speech_nrg, 1 ); cannam@154: } cannam@154: /* Power scaling */ cannam@154: if( speech_nrg <= 0 ) { cannam@154: SA_Q15 = silk_RSHIFT( SA_Q15, 1 ); cannam@154: } else if( speech_nrg < 16384 ) { cannam@154: speech_nrg = silk_LSHIFT32( speech_nrg, 16 ); cannam@154: cannam@154: /* square-root */ cannam@154: speech_nrg = silk_SQRT_APPROX( speech_nrg ); cannam@154: SA_Q15 = silk_SMULWB( 32768 + speech_nrg, SA_Q15 ); cannam@154: } cannam@154: cannam@154: /* Copy the resulting speech activity in Q8 */ cannam@154: psEncC->speech_activity_Q8 = silk_min_int( silk_RSHIFT( SA_Q15, 7 ), silk_uint8_MAX ); cannam@154: cannam@154: /***********************************/ cannam@154: /* Energy Level and SNR estimation */ cannam@154: /***********************************/ cannam@154: /* Smoothing coefficient */ cannam@154: smooth_coef_Q16 = silk_SMULWB( VAD_SNR_SMOOTH_COEF_Q18, silk_SMULWB( (opus_int32)SA_Q15, SA_Q15 ) ); cannam@154: cannam@154: if( psEncC->frame_length == 10 * psEncC->fs_kHz ) { cannam@154: smooth_coef_Q16 >>= 1; cannam@154: } cannam@154: cannam@154: for( b = 0; b < VAD_N_BANDS; b++ ) { cannam@154: /* compute smoothed energy-to-noise ratio per band */ cannam@154: psSilk_VAD->NrgRatioSmth_Q8[ b ] = silk_SMLAWB( psSilk_VAD->NrgRatioSmth_Q8[ b ], cannam@154: NrgToNoiseRatio_Q8[ b ] - psSilk_VAD->NrgRatioSmth_Q8[ b ], smooth_coef_Q16 ); cannam@154: cannam@154: /* signal to noise ratio in dB per band */ cannam@154: SNR_Q7 = 3 * ( silk_lin2log( psSilk_VAD->NrgRatioSmth_Q8[b] ) - 8 * 128 ); cannam@154: /* quality = sigmoid( 0.25 * ( SNR_dB - 16 ) ); */ cannam@154: psEncC->input_quality_bands_Q15[ b ] = silk_sigm_Q15( silk_RSHIFT( SNR_Q7 - 16 * 128, 4 ) ); cannam@154: } cannam@154: cannam@154: RESTORE_STACK; cannam@154: return( ret ); cannam@154: } cannam@154: cannam@154: /**************************/ cannam@154: /* Noise level estimation */ cannam@154: /**************************/ cannam@154: # if !defined(OPUS_X86_MAY_HAVE_SSE4_1) cannam@154: static OPUS_INLINE cannam@154: #endif cannam@154: void silk_VAD_GetNoiseLevels( cannam@154: const opus_int32 pX[ VAD_N_BANDS ], /* I subband energies */ cannam@154: silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */ cannam@154: ) cannam@154: { cannam@154: opus_int k; cannam@154: opus_int32 nl, nrg, inv_nrg; cannam@154: opus_int coef, min_coef; cannam@154: cannam@154: /* Initially faster smoothing */ cannam@154: if( psSilk_VAD->counter < 1000 ) { /* 1000 = 20 sec */ cannam@154: min_coef = silk_DIV32_16( silk_int16_MAX, silk_RSHIFT( psSilk_VAD->counter, 4 ) + 1 ); cannam@154: /* Increment frame counter */ cannam@154: psSilk_VAD->counter++; cannam@154: } else { cannam@154: min_coef = 0; cannam@154: } cannam@154: cannam@154: for( k = 0; k < VAD_N_BANDS; k++ ) { cannam@154: /* Get old noise level estimate for current band */ cannam@154: nl = psSilk_VAD->NL[ k ]; cannam@154: silk_assert( nl >= 0 ); cannam@154: cannam@154: /* Add bias */ cannam@154: nrg = silk_ADD_POS_SAT32( pX[ k ], psSilk_VAD->NoiseLevelBias[ k ] ); cannam@154: silk_assert( nrg > 0 ); cannam@154: cannam@154: /* Invert energies */ cannam@154: inv_nrg = silk_DIV32( silk_int32_MAX, nrg ); cannam@154: silk_assert( inv_nrg >= 0 ); cannam@154: cannam@154: /* Less update when subband energy is high */ cannam@154: if( nrg > silk_LSHIFT( nl, 3 ) ) { cannam@154: coef = VAD_NOISE_LEVEL_SMOOTH_COEF_Q16 >> 3; cannam@154: } else if( nrg < nl ) { cannam@154: coef = VAD_NOISE_LEVEL_SMOOTH_COEF_Q16; cannam@154: } else { cannam@154: coef = silk_SMULWB( silk_SMULWW( inv_nrg, nl ), VAD_NOISE_LEVEL_SMOOTH_COEF_Q16 << 1 ); cannam@154: } cannam@154: cannam@154: /* Initially faster smoothing */ cannam@154: coef = silk_max_int( coef, min_coef ); cannam@154: cannam@154: /* Smooth inverse energies */ cannam@154: psSilk_VAD->inv_NL[ k ] = silk_SMLAWB( psSilk_VAD->inv_NL[ k ], inv_nrg - psSilk_VAD->inv_NL[ k ], coef ); cannam@154: silk_assert( psSilk_VAD->inv_NL[ k ] >= 0 ); cannam@154: cannam@154: /* Compute noise level by inverting again */ cannam@154: nl = silk_DIV32( silk_int32_MAX, psSilk_VAD->inv_NL[ k ] ); cannam@154: silk_assert( nl >= 0 ); cannam@154: cannam@154: /* Limit noise levels (guarantee 7 bits of head room) */ cannam@154: nl = silk_min( nl, 0x00FFFFFF ); cannam@154: cannam@154: /* Store as part of state */ cannam@154: psSilk_VAD->NL[ k ] = nl; cannam@154: } cannam@154: }