FreeCalypso > hg > gsm-codec-lib
view libgsmefr/levinson.c @ 242:f081a6850fb5
libgsmfrp: new refined implementation
The previous implementation exhibited the following defects,
which are now fixed:
1) The last received valid SID was cached forever for the purpose of
handling future invalid SIDs - we could have received some valid
SID ages ago, then lots of speech or NO_DATA, and if we then get
an invalid SID, we would resurrect the last valid SID from ancient
history - a bad design. In our new design, we handle invalid SID
based on the current state, much like BFI.
2) GSM 06.11 spec says clearly that after the second lost SID
(received BFI=1 && TAF=1 in CN state) we need to gradually decrease
the output level, rather than jump directly to emitting silence
frames - we previously failed to implement such logic.
3) Per GSM 06.12 section 5.2, Xmaxc should be the same in all 4 subframes
in a SID frame. What should we do if we receive an otherwise valid
SID frame with different Xmaxc? Our previous approach would
replicate this Xmaxc oddity in every subsequent generated CN frame,
which is rather bad. In our new design, the very first CN frame
(which can be seen as a transformation of the SID frame itself)
retains the original 4 distinct Xmaxc, but all subsequent CN frames
are based on the Xmaxc from the last subframe of the most recent SID.
author | Mychaela Falconia <falcon@freecalypso.org> |
---|---|
date | Tue, 09 May 2023 05:16:31 +0000 |
parents | 1cdbaeec7bcc |
children |
line wrap: on
line source
/************************************************************************* * * FUNCTION: Levinson() * * PURPOSE: Levinson-Durbin algorithm in double precision. To compute the * LP filter parameters from the speech autocorrelations. * * DESCRIPTION: * R[i] autocorrelations. * A[i] filter coefficients. * K reflection coefficients. * Alpha prediction gain. * * Initialisation: * A[0] = 1 * K = -R[1]/R[0] * A[1] = K * Alpha = R[0] * (1-K**2] * * Do for i = 2 to M * * S = SUM ( R[j]*A[i-j] ,j=1,i-1 ) + R[i] * * K = -S / Alpha * * An[j] = A[j] + K*A[i-j] for j=1 to i-1 * where An[i] = new A[i] * An[i]=K * * Alpha=Alpha * (1-K**2) * * END * *************************************************************************/ #include "gsm_efr.h" #include "typedef.h" #include "namespace.h" #include "basic_op.h" #include "oper_32b.h" #include "no_count.h" #include "sig_proc.h" #include "cnst.h" #include "enc_state.h" /* Lpc order == 10 */ #define M 10 void Levinson ( struct EFR_encoder_state *st, Word16 Rh[], /* (i) : Rh[m+1] Vector of autocorrelations (msb) */ Word16 Rl[], /* (i) : Rl[m+1] Vector of autocorrelations (lsb) */ Word16 A[], /* (o) : A[m] LPC coefficients (m = 10) */ Word16 rc[] /* (o) : rc[4] First 4 reflection coefficients */ ) { Word16 i, j; Word16 hi, lo; Word16 Kh, Kl; /* reflexion coefficient; hi and lo */ Word16 alp_h, alp_l, alp_exp; /* Prediction gain; hi lo and exponent */ Word16 Ah[M + 1], Al[M + 1]; /* LPC coef. in double prec. */ Word16 Anh[M + 1], Anl[M + 1];/* LPC coef.for next iteration in double prec. */ Word32 t0, t1, t2; /* temporary variable */ /* K = A[1] = -R[1] / R[0] */ t1 = L_Comp (Rh[1], Rl[1]); t2 = L_abs (t1); /* abs R[1] */ t0 = Div_32 (t2, Rh[0], Rl[0]); /* R[1]/R[0] */ test (); if (t1 > 0) t0 = L_negate (t0); /* -R[1]/R[0] */ L_Extract (t0, &Kh, &Kl); /* K in DPF */ rc[0] = round (t0); move16 (); t0 = L_shr (t0, 4); /* A[1] in */ L_Extract (t0, &Ah[1], &Al[1]); /* A[1] in DPF */ /* Alpha = R[0] * (1-K**2) */ t0 = Mpy_32 (Kh, Kl, Kh, Kl); /* K*K */ t0 = L_abs (t0); /* Some case <0 !! */ t0 = L_sub ((Word32) 0x7fffffffL, t0); /* 1 - K*K */ L_Extract (t0, &hi, &lo); /* DPF format */ t0 = Mpy_32 (Rh[0], Rl[0], hi, lo); /* Alpha in */ /* Normalize Alpha */ alp_exp = norm_l (t0); t0 = L_shl (t0, alp_exp); L_Extract (t0, &alp_h, &alp_l); /* DPF format */ /*--------------------------------------* * ITERATIONS I=2 to M * *--------------------------------------*/ for (i = 2; i <= M; i++) { /* t0 = SUM ( R[j]*A[i-j] ,j=1,i-1 ) + R[i] */ t0 = 0; move32 (); for (j = 1; j < i; j++) { t0 = L_add (t0, Mpy_32 (Rh[j], Rl[j], Ah[i - j], Al[i - j])); } t0 = L_shl (t0, 4); t1 = L_Comp (Rh[i], Rl[i]); t0 = L_add (t0, t1); /* add R[i] */ /* K = -t0 / Alpha */ t1 = L_abs (t0); t2 = Div_32 (t1, alp_h, alp_l); /* abs(t0)/Alpha */ test (); if (t0 > 0) t2 = L_negate (t2); /* K =-t0/Alpha */ t2 = L_shl (t2, alp_exp); /* denormalize; compare to Alpha */ L_Extract (t2, &Kh, &Kl); /* K in DPF */ test (); if (i < 5) { rc[i - 1] = round (t2); move16 (); } /* Test for unstable filter. If unstable keep old A(z) */ if (abs_s (Kh) > 32750) { for (j = 0; j <= M; j++) { A[j] = st->old_A[j]; } for (j = 0; j < 4; j++) { rc[j] = 0; } return; } /*------------------------------------------* * Compute new LPC coeff. -> An[i] * * An[j]= A[j] + K*A[i-j] , j=1 to i-1 * * An[i]= K * *------------------------------------------*/ for (j = 1; j < i; j++) { t0 = Mpy_32 (Kh, Kl, Ah[i - j], Al[i - j]); t0 = L_mac (t0, Ah[j], 32767); t0 = L_mac (t0, Al[j], 1); L_Extract (t0, &Anh[j], &Anl[j]); } t2 = L_shr (t2, 4); L_Extract (t2, &Anh[i], &Anl[i]); /* Alpha = Alpha * (1-K**2) */ t0 = Mpy_32 (Kh, Kl, Kh, Kl); /* K*K */ t0 = L_abs (t0); /* Some case <0 !! */ t0 = L_sub ((Word32) 0x7fffffffL, t0); /* 1 - K*K */ L_Extract (t0, &hi, &lo); /* DPF format */ t0 = Mpy_32 (alp_h, alp_l, hi, lo); /* Normalize Alpha */ j = norm_l (t0); t0 = L_shl (t0, j); L_Extract (t0, &alp_h, &alp_l); /* DPF format */ alp_exp = add (alp_exp, j); /* Add normalization to alp_exp */ /* A[j] = An[j] */ for (j = 1; j <= i; j++) { Ah[j] = Anh[j]; move16 (); Al[j] = Anl[j]; move16 (); } } A[0] = 4096; move16 (); for (i = 1; i <= M; i++) { t0 = L_Comp (Ah[i], Al[i]); st->old_A[i] = A[i] = round (L_shl (t0, 1)); } return; }