AMR-NB-WIP

From MultimediaWiki
Revision as of 05:03, 8 September 2007 by Superdump (talk | contribs) (Initial import of my notes on AMR NB into a temporary page until they are migrated to the main AMR NB page)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

AMR narrow band decoder

This text aims to be a simpler and more explicit document of the AMR narrow band decoding processes to aid in development of a decoder. Reference to sections of the specification will be made in the following format: (c.f. 5.2.5). Happy reading.

Nomenclature weirdness

Throughout the specification, a number of references are made to the same (or very similar) items with fairly confusing variation. They are listed below to aid understanding of the following text but efforts will be made to consistently use one item name throughout or to use both with the lesser used name in parenthesis.

  • Pitch / Adaptive codebook
  • Fixed / Innovative (also algebraic when referring to the codebook)
  • Quantified means not quantised


Summary

  • Mode dependent bitstream parsing
  • Indices parsed from bitstream
  • Indices decoded to give LSF vectors, fractional pitch lags, innovative code vectors and the pitch and innovative gains
  • LSF vectors converted to LP filter coefficients at each subframe
  • Subframe decoding
    • Excitation vector = adaptive code vector * adaptive (pitch) gain + innovative code vector * innovative gain
    • Excitation vector filtered through an LP synthesis filter to reconstruct speech
    • Speech signal filtered with adaptive postfilter


Bitstream parsing

Documented on http://wiki.multimedia.cx/index.php?title=AMR-NB and in 26.101 For implementation, see http://svn.mplayerhq.hu/soc/amr/amrnbdec.c?view=markup


Decoding of LP filter parameters

The received indices of LSP quantization are used to reconstruct the quantified LSP vectors. (c.f. 5.2.5)

12.2 - 5 split matrix quantised LSF indices to 2 LSP vectors

  • indices into code books are parsed from the bit stream
  • indices give elements of split matrix quantised (SMQ) residual LSF vectors from the relevant code books
  • prediction from the previous frame is added to obtain the mean-removed LSF vectors
  • the mean is added
  • the LSF vectors are converted to cosine domain LSP vectors


Indices give elements of split matrix quantised (SMQ) residual LSF vectors from the relevant code books

The elements of the SMQ vectors are stored at an index into a code book that varies according to the mode. There are 5 code books for the 12.2kbps mode corresponding to the 5 indices. These tables will be referred to as:

lsf_m_n

where m is the number of indices parsed according to the mode n is the index 'position' i.e. 1 for the first index, etc

The 5 indices are stored using 7, 8, 8 + sign bit, 8, 6 bits respectively. The four elements of a 'split quantized sub-matrix' are stored at the index position in the appropriate code book are:

1st index in 1st code book : r1_1, r1_2, r2_1, r2_2; 2nd index in 2nd code book : r1_3, r1_4, r2_3, r2_4; ...

with rj_i where j indicates the first or second residual lsf vector and i indicates the coefficient of a residual lsf vector ( i = 1, ..., 10 )

rj_i - residual line spectral frequencies (LSFs) in Hz


  • prediction from the previous frame is added to obtain the mean-removed LSF
 vectors

zj(n) = rj(n) + 0.65*^r2(n-1)

where zj(n) is a mean-removed LSF vector from the current frame (denoted n) ^r2(n-1) is the quantified 2nd residual vector of the last frame (denoted n-1)


  • the mean is added

fj = zj + lsf_mean_m

with lsf_mean_m is a table of the means of the lsf coefficients where m is the number of indices parsed according to the mode


  • the LSF vectors are converted to cosine domain LSP vectors

qk_i = cos( fj_i * 2 * π / f_s )

qk_i - line spectral pairs (LSPs) in the cosine domain k = 2*j (the two lsf vecs give the LSP vecs q2, q4 at the 2nd and 4th subframes) fj_i are in [0,4000] Hz f_s - sampling frequency in Hz (8kHz)


Other active modes - 3 sub vectors to 1 LSP vector

The process for the other modes is similar to that for the 12.2kbps mode.

  • indices into code books are parsed from the bit stream
  • indices give elements of a split matrix quantised (SMQ) residual LSF vector
 from the relevant code books
  • prediction from the previous frame is added to obtain the mean-removed LSF
 vector
  • the mean is added
  • the LSF vector is converted to a cosine domain LSP vector


  • indices give elements of a split matrix quantised (SMQ) residual LSF vector
 from the relevant code books

The 3 indices are stored with the following numbers of bits:

10.2 kbit/s 8 9 9 7.95 kbit/s 9 9 9 7.40 kbit/s 8 9 9 6.70 kbit/s 8 9 9 5.90 kbit/s 8 9 9 5.15 kbit/s 8 8 7 4.75 kbit/s 8 8 7

The four elements of a 'split quantized sub-matrix' are stored at the index position in the appropriate code book are:

1st index in 1st code book : r_1, r_2, r_3; 2nd index in 2nd code book : r_4, r_5, r_6; 3rd index in 3rd code book : r_7, r_8, r_9, r_10;

with residual lsf vector r_i where i indicates the coefficient of vector ( i = 1, ..., 10 )

r_i - residual line spectral frequencies (LSFs) in Hz


  • prediction from the previous frame is added to obtain the mean-removed LSF
 vector

z_i(n) = r_i(n) + pred_fac_i * ^r_i(n-1)

where z_i(n) is the mean-removed LSF vector from the current frame (denoted n) pred_fac_i is the prediction factor for the ith LSF coefficient ^r_i(n-1) is the quantified residual vector of the last frame (denoted n-1)

These processes give the LSP vector at the 4th subframe (q4)


The available LSP vector(s) are used to linearly interpolate vectors for the other subframes. 5.2.6

12.2 q1(n) = 0.5*q4(n-1) + 0.5*q2(n) q3(n) = 0.5*q2(n) + 0.5*q4(n)

Others q1(n) = 0.75*q4(n-1) + 0.25*q4(n) q2(n) = 0.5 *q4(n-1) + 0.5 *q4(n) q3(n) = 0.25*q4(n-1) + 0.75*q4(n)


The LSP vector is converted to LP filter coefficients. 5.2.4

for i=1..5

 f1_i  = 2*f1(i-2) - 2 * q_2i-1 * f1(i-1)
 for j=i-1..1
   f1_j +=   f1(j-2) - 2 * q_2i-1 * f1(j-1)
 end

end

f1_-1 = 0; f1_0 = 0;

Same for f2_i with q_2i insteand of q_2i-1

for i=1..5

 f'1_i = f1_i + f1_i-1
 f'2_i = f2_i - f2_i-1

end

for i=1..5

 a_i = 0.5*f'1_i    + 0.5*f'2_i

end for i=6..10

 a_i = 0.5*f'1_11-i - 0.5*f'2_11-i

end

a_i are the LP filter coefficients



Decoding of the adaptive (pitch) codebook vector ................................................

  • indices parsed from bitstream
  • indices give integer and fractional parts of the pitch lag
  • adaptive codebook vector v(n) is found by interpolating the past excitation
 u(n) at the pitch lag using an FIR filter. 5.6

12.2kbps mode - 1/6 resolution pitch lag

  • indices give integer and fractional parts of the pitch lag

In the first and third subframes, a fractional pitch delay is used with resolutions: 1/6 in the range [17 3/6,94 3/6] and integers only in the range [95, 143].

The lower bound of the pitch lag is 17.5 and the fractional part is in 1/6 resolution, so the pitch index is given by:

pitch index = integer part*6 + fractional part -17.5*6

so theoretically...

if(pitch_index < (94 4/6 - 17 3/6)*6)

 // fractional part is encoded
 pitch_lag_int  = pitch_index/6 + 17 3/6;
 pitch_lag_frac = pitch_index - pitch_lag_int*6 + (17 3/6)*6;
 but the reference source adds an extra 2/6, i assume for rounding:
 pitch_lag_int = (pitch_index + 5)/6 + 17;

else

 // only integer part encoded, no fractional part
 pitch_lag_int  = pitch_index - 368;
 pitch_lag_frac = 0;

I have not yet discovered the meaning of 368 (368/6 = 61 2/6)

For the second and fourth subframes, a pitch resolution of 1/6 is always used in the range [T1−5 3/6,T1+4 3/6], where T1 is nearest integer to the fractional pitch lag of the previous (1st or 3rd) subframe, bounded by 18...143.

// find the search range search_range_min = max(pitch_lag_int - 5, 18); (only subtract 5 because of the above mentioned rounding?)

search_range_max = search_range_min + 9; if(search_range_max > 143) {

 search_range_max = 143;
 search_range_min = search_range_max - 9;

} (only add/subtract 9 because of the above mentioned rounding?)

// calculate the pitch lag pitch_lag_int = (pitch_index + 5) + search_range_min - 1; pitch_lag_frac = -2;

The pitch delay is encoded with 9 bits in the first and third subframes and the relative delay of the other subframes is encoded with 6 bits.


Decoding of the algebraic (innovative or fixed) codebook vector ...............................................................

The parsed algebraic codebook index is used to find the positions and amplitudes (signs) of the excitation pulses and to find the algebraic codebook vector c(n).

If the integer part of the pitch lag, T, is less than the subframe size 40, the pitch sharpening procedure is applied which translates into c(n) += βc(n−T) , where β is the decoded pitch gain, ^g_p, bounded by [0.0,1.0] or [0.0,0.8], depending on mode.



Decoding of the adaptive and fixed codebook gains .................................................

12.2kbps and 7.95kbps - scalar quantised gains: The received indices are used to find the adaptive codebook gain, ^g_p, and the algebraic codebook gc factor, ^γ_gc (gc for gain correction), from the corresponding quantisation tables.

Other modes - vector quantised gains: The received index gives both the adaptive codebook gain, ^g_p, and the algebraic codebook gc factor, ^γ_gc. The estimated algebraic codebook gain gc′ is found as described in clause 5.7.



Smoothing of the fixed codebook gain ....................................

10.2, 6.70, 5.90, 5.15, 4.75 kbit/s modes

Adaptive smoothing of fixed codebook gain. 6.1 4)



Anti-sparseness processing ..........................

7.95, 6.70, 5.90, 5.15, 4.75 kbit/s modes

An adaptive anti-sparseness postprocessing procedure is applied to the fixed codebook vector c(n) in order to reduce perceptual artifacts arising from the sparseness of the algebraic fixed codebook vectors with only a few non-zero samples per subframe. The anti-sparseness processing consists of circular convolution of the fixed codebook vector with an impulse response. Three pre-stored impulse responses are used and a number impNr = 0,1,2 is set to select one of them. A value of 2 corresponds to no modification, a value of 1 corresponds to medium modification, while a value of 0 corresponds to strong modification. The selection of the impulse response is performed adaptively from the adaptive and fixed codebook gains. See 6.1 5)



Computing the reconstructed speech ..................................

Construct excitation:

u(n) = ^g_p.v(n) + ^g_c.c(n)

6.1 6)



Additional instability protection .................................

6.1 7)



Adaptive post-filtering .......................

6.2.1



High-pass filtering and upscaling .................................

6.2.2