comment about reduced codelength component in LTP quantizer error metric
authorKoen Vos <koenvos@users.noreply.github.com>
Thu, 2 Jun 2016 15:03:27 +0000 (23:03 +0800)
committerJean-Marc Valin <jmvalin@jmvalin.ca>
Tue, 19 Jul 2016 20:11:19 +0000 (16:11 -0400)
silk/VQ_WMat_EC.c

index 2f3aa82..8b5ebde 100644 (file)
@@ -65,7 +65,7 @@ void silk_VQ_WMat_EC_c(
     *ind = 0;
     for( k = 0; k < L; k++ ) {
         /* Weighted rate */
-        /* Quantization error: 1 - 2* xX * cb + cb' * XX * cb */
+        /* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */
         sum1_Q15 = SILK_FIX_CONST( 1.001, 15 );
 
         /* first row of XX_Q17 */
@@ -107,7 +107,8 @@ void silk_VQ_WMat_EC_c(
         if( sum1_Q15 >= 0 ) {
             /* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */
             bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 ) - (15 << 7) );
-            bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 2 );
+            /* In the following line we reduce the codelength component by half ("-1"); seems to slghtly improve quality */
+            bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 );
             if( bits_tot_Q8 <= *rate_dist_Q8 ) {
                 *rate_dist_Q8 = bits_tot_Q8;
                 *res_nrg_Q15 = sum1_Q15;