Implementation of G.729 algorithm on TI

1
Implementation of G.729 algorithm on TIs
TMS320C62xx
Zoha Pajouhi
2
Outline

• What is G.729?
• How Does it work ?
• ????Acquaintance with C62xx processors
• Steps to be taken for implementation
• Optimizations required for implementation
• Implementation of the algorithm
• Summary Conclusions

3
What is G.729 ?

• G.729 is a speech coding technique
• Most important usage is in VoIP
• Compress speech signal from 64kbps to 8kbps
• It uses the CS-ACELP algorithm
• CS-ACELP stands for conjugate structure algebraic
  code excited linear prediction
• Has different annexes G.729 Annex A , G.729
  Annex B , G.729 Annex D , G.729 Annex E .
• Different in difficulty of the algorithm
  ,utilizing voice activity detection , reduced bit
  rates at the cost of lower quality vice versa
• I discuss implementing G.729 Annex A

4
How Does G.729 Work ? (1)

• Uses input frames of 10ms which is equal to 80
  samples
• Each frame has two subframes of 5ms
• The idea of the algorithm is to predict the next
  coming signals by means of linear prediction .
• It also uses statistical data to distinguish the
  resemblance of the signal to special signals in
  its code book
• Exact signal variation is irrelevant to our
  discussion .

5
How Does G.729 Work ? (2)

• These operations are performed once per frame
• Pre processing scales down the signal by a
  factor of 2 , passing from a HP filter.
• LP (linear prediction) analysis Uses linear
  prediction to model the signal , the LP
  coefficients are converted to LPC coefficients
  for less sensitivity to quantization noise .

6
How Does G.729 Work ? (3)

1. Quantization LPCs are quantized and used
   throughout the rest of the algorithm .
2. Open-loop pitch analysis pitch analysis is too
   difficult , this part gives us a rough estimation
   of the pitch

7
How Does G.729 Work ? (4)

• These operations are performed twice a frame or
  once per subframe
• Closed-loop pitch analysis Determines exact
  pitch delay through a closed loop
• Fixed codebook search computes the resemblance
  of the signal to the different codes in the
  codebook is used for consonant sounds
• Adaptive codebook search Is the same as fixed
  codebook except for it is for nonconsonant sounds
  .

8
How Does G.729 Work ? (5)

• The bit allocation is as follows

9
Why use DSP processors ?

• Complex long calculations in speech coding
  algorithms
• Short design period
• Lead to using DSP processors instead of custom
  hardware design
• Different DSP processors available floating and
  fixed point processors usually trade off in price
  vs. precision.

10
DSP processor
11
Acquaintance with DSP C62xx processors

• 150-250 MHz clock
• 8 instructions/clk
• 1200-2000 million instructions/s (MIPs)
• 8 , 16 32 bit data manipulation better memory
  access
• Low power
• VLIW code structure ,6 ALUs ,2 multipliers
• 40-bit mathematic function
• Saturation and normalization blocks
• On-chip Ram , etc.

12
Steps to be taken for implementation

• System simulation analysis Matlab or C or C
  is usually used .
• Simulation of the implementation on the processor
  one can make different optimization choices
  according to the processor being used .
• Optimization will be discussed later
• Conversion to processor assembly

13
Optimization

• Different issues should be considered
• Processor independent issues
• Locating independent instructions between two
  dependant ones to use parallelism
• Optimized usage of registers
• Loop folding/unfolding .
• Folding for reduced code memory
• Unfolding for reducing loop overhead
• Using pointers instead of arrays

14
Optimization

• Processor dependant issues
• Substituting special functions instead of
  assembly e.g. saturation has got a special
  instruction doesnt need to be rewritten .
• Reading 32 bit data instead of 16 bit in 16 bit
  operations to reduce memory access time
• Pipelining the algorithm sometimes done by the
  design software sometimes not e.g. the inner
  loops are pipelined but not the outer ones.
• Using hardware over flow bit , not supported in
  C6000 series .

15
Implementation of the algorithm

• The multichannel system can run more than one
  algorithm at the same time.
• Any algorithm compliant with the eXpress DSP
  Algorithm Standard (xDAIS) is capable of
  multichannel processing.
• An xDAIS-compliant algorithm requires three
  functional modules
• initialization, freeing and kernel.
• The kernel module performs the algorithm
  processing while
• the initialization and freeing module
  initializes/frees the algorithm context data.

16
Framework initialization

• Initialize context data and store to the desired
  memory location.
• The system repeatedly calls an algorithm/algorithm
  s until all the frames have been processed

17
Algorithm implementation

• The G.729 speech encoder is divided into five
  submodules
• Submodule 1 contains pre-processing, the LP
  analysis and LPC to LSP conversion ,including
  Pre_Process(), Autocorr(), Lag_window(),
  Levinson() and Az_lsp().
• Submodule 2 calls Qua_lsp() to conduct LSP
  quantization.
• Submodule 3 generates the interpolated LPC
  parameters, computes weighted speech, and finds
  the open-loop pitch, including Int_qlpc(),
  Int_lpc(), Weight_Az(), Residu(), Syn_filt() and
  Pitch_ol().
• Submodules 1 through 3 are for frame processing
  and should be done once per frame. Submodules

18
Algorithm Implementation

• Submodule 4 performs the closed-loop fractional
  pitch search and the adaptive codebook search,
  calling Pitch_fr3(), Enc_lag3(), Pred_lt_3(),
  Convolve() and G_pitch().
• Submodule 5 performs the innovative codebook
  search and filter memory update, calling
  ACELP_Codebook(), Corr_xy2(), Qua_gain() and
  Syn_filt().
• 4 and 5 are for subframe processing and should be
  repeated twice per frame

19
Data Memory requirements

• The data memory is divided into three groups
• Context data The context data are the static
  variables and arrays with values that must be
  kept from one frame to the next.
• Tables The constant tables are sorted into the
  G729_TABLES ,these tables contain different
  constants needed for the algorithm
• Local variables and arrays The local variables
  and arrays are stored in the stacks to be simply
  used when the algorithm is implemented .

20
Summary Conclusion

• The G.729 standard is a popular choice for
  applications, such as VoIP, that require
  efficient use of bandwidth and good speech
  quality.
• This standard has a good balance of bit-rate and
  frame size, producing acceptable speech quality
• The TI DSP processors are capable of performing
  the algorithm
• Although implementation seems simple but there
  are various issues to be considered .

21
References

1. ITU-T G.729 Annex AReduced Complexity 8 kb/s
   CS-ACELP Codec for Digital Simultaneous Voice and
   Data , Redwan Salami, Claude Laflamme, Bruno
   Bessette, and Jean-Pierre AdoulUniversity of
   Sherbrooke , IEEE Communications Magazine
   ,September 1997
2. CODING OF SPEECH AT 8 kbit/s USING
   CONJUGATE-STRUCTURE ALGEBRAIC-CODE-EXCITED
   LINEAR-PREDICTION (CS-ACELP) , ITU-T
   Recommendation G.729 , (03/96)
3. G.729/A Speech Coder Multichannel TMS320C62x
   Implementation , Chiouguey Chen ,Xiangdong Fu ,
   TI Application Report SPRA564B - February 2000
4. TMS320C6211, TMS320C6211BFIXED-POINT DIGITAL
   SIGNAL PROCESSORS ,AUGUST 1998 ,REVISED MARCH
   2004 ,TI corp. Datasheets