Research Experience

1
Research Experience

• Wang Jianhong
• CECS Department
• University of Missouri-Columbia
• 201 Engineering Building West
• Columbia, MO 65211
• Phone [phone]
• Email jwcdb_at_missouri.edu

2
Digital Acoustic Echo Cancellation(AEC) Algorithm

• Theoretical basis of echo cancellation is
  adaptive filter. There are two basic categories
  of algorithms for acoustic echo cancellation.
• Category One LMS(Least Mean Squares) Algorithm
  The criterion function is the expected value of
  the squared error and the coefficients of the
  adaptive filter are updated according to the
  stochastic steepest descent algorithm. This
  algorithm is widely used due to its comparatively
  lease computational requirements and its robust
  stability characteristics. But the convergence
  slows when the input signals are none-stationary
  or the adaptive filters have a large number of
  taps.

3
Digital Acoustic Echo Cancellation(AEC) Algorithm

• Category Two LS(Least Squares) Algorithm The
  criterion is to minimize the squared error summed
  over time. The advantage of the algorithm is the
  fast convergence with a penalty of the increased
  computational requirement. And this algorithm is
  prone to numeric instability which impedes its
  implementation on the fixed point digital signal
  processor. FTF is a fast algorithm in this
  category.
• Mixed FTF-NLMS Algorithm LS algorithm is always
  stable in the starting period before all the
  filter coefficients are updated by a non-zero
  value. The disastrous overflow happens after the
  starting period. So a new adaptive method is
  built that in the starting period the LS
  algorithm is used while after this period the LMS
  algorithm is used.

4
Digital Acoustic Echo Cancellation(AEC) Algorithm

• Brief Description of Mixed FTF-NLMS Algorithm
• 1. Initialization Set the taps of the adaptive
  filter zero.
• 2. During the starting period, using FTF
  algorithm to do the acoustic echo cancellation.
• 3. Then using NLMS algorithm. In this stage, a
  monitor is embedded in the filter. If the effect
  of the echo cancellation is not good. The
  adaptive filter will be switched to FTF
  algorithm.

5
Digital Acoustic Echo Cancellation(AEC)
AlgorithmSimulation Result Comparison between
NLMS and proposed algorithm.
6
Digital Acoustic Echo Cancellation(AEC)
AlgorithmSimulation Result Stability
characteristics of the proposed algorithm.
7
The Optimal Lattice Architectures for the
Real-Time Computation of DFT/IDFT

• IntroductionIn real-time signal processing
  application such as COFDM, DFT/IDFT is an
  important module. Although we can take advantage
  of fast algorithm FFT/IFFT, it still makes the
  VLSI implementation very cost expensive when we
  have to deal with a large block size of FFT/IFFT.
  Several efficient schemes to VLSI implementation
  will be proposed below.

8
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT

• One of the schemes which based on time-recursive
  MDCT/MDST. We first decompose the IFFT/FFT into a
  combination of a DCT-like and a DST-like
  transform. By exploiting the symmetric property
  of the Fourier transform as well as the fast
  algorithms of the DCT and DST, we derive a new
  scheme which only involve real-valued operators
  to replace the complex-domain IFFT/FFT. The
  structure computes the transformed data from
  sequential inputs in a time-recursive way and
  only requires O(N) hardware complexity.

9
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT

• The IFFT/FFT block diagram in this new scheme is
  showed below.

10
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT

• At the transmitter side, to ensure the IFFT
  generates only real-valued outputs, the inputs of
  the IFFT have do some changes
• The IFFT of a data sequence of length 2N is

11
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT

• By applying the conjugate-symmetric property of
  the input data, we can get
• The computation of the IFFT is decomposed into
  two part of real-valued operators. One is a
  discrete cosine transform like operation. The
  other is a discrete sine transform like
  operation. we will term these two transform as
  modified DCT (MDCT) and DST (MDST).

12
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT

• Note that MDCT and MDST involve only real-valued
  computation. This will reduce the computational
  complexity drastically.
• We employ the time-recursive approach to
  implement the IFFT module. It is an efficient
  implementation of the dual generation of the MDCT
  and MDST .
• We define the MDCT of a sequential input data
  starting from x(t) and ending with x(tN-1) as
  follows. After the new datum arrives , the MDCT
  is updated as

13
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT
14
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT

• For block processing, the initial states are
  zeros.

15
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT According to
the equations above, we can derive the complete
time-recursive lattice module as shown in the
figure below
16
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT

• The implementation of the MDST is similar with
  the method of MDCT.
• Complexity Analysis
• In the direct implementation using the butterfly
  structure , it consists log2(2N) stages in 2N
  points FFT/IFFT. Each stage consists of N
  multipliers and 2N adders. Note that the input
  data are complex data and it requires 4
  real-valued multipliers and 2 real-valued adders
  for 1 complex multiplier also it requires 2
  real-valued adders for 1 complex adder. Hence,
  the direct approach requires a total of 4N
  log2(2N) real-value multipliers and 6Nlog2(2N)
  real adders.

17
The Optimal Lattice Architectures for the
Real-Time Computation of FFT/IFFT

• In our approach, we decompose the IFFT into the
  real-valued transform kernels (MDCT and MDST). As
  a result, we only need real operations to realize
  IFFT structure. Therefore, we only need a total
  of 4(N-1) real multipliers and 5(N-1)2
  real-valued adders.
• For the FFT module, we need one SC which is an
  accumulator and (N-1) lattice modules (Mn) The
  overall architecture of the FFT needs a total of
  4(N-1) real multipliers and 3(N-1)1 real adders.
• If N is large ,the computational complexity of
  our scheme is much smaller than conventional
  direct approach.

18
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Introduction More and More high-quality, digital
  stereo media is introduced to the public,
  including the CD, digital audio tape (DAT),
  mini-disk (MD), and digital compact cassette
  (DCC). All these media assume a 20-kHz bandwidth
  for the audio signal. The CD and DAT use a 16-bit
  pulse-code modulation format and sampling rates
  of 44.1 and 48 kHz. Thus bit rates of 1.406 and
  1.536 Mbits/s are used, respectively, for
  two-channel stereo. In 1992, the international
  standard ISO/IEC 11172 Coding of Moving Pictures
  and Associated Audio for Digital Storage Media at
  up to about 1.5 Mbit/s was finalized, which is
  also known as the MPEG-1 standard. MPEG audio
  about Hi-Fi audio coding offers a choice of three
  independent layers of compression.

19
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Introduction Layer III is the most complex but
  offer the best audio quality, particularly for
  bit rates around 64 Kbps per channel. This layer
  suits audio transmission over ISDN and Internet.
  A MPEG audio layer III codec on a general purpose
  DSP chip TMS320C54 is presented below.
• The MPEG audio layer III codec is a complicated
  compound of digital signal processing algorithms
  ranging from subband transform to huffman coding.
  
• Due to the comparably small computational
  requirements of decoder part, only a small
  general purpose DSP chip will be required.

20
Developing a MPEG Audio Layer III Codec on
TMS320C549

• The encoder part is much more complicated and the
  computational demands required to estimate the
  parameters of psycho-acoustic model and operate
  the noise allocation are very huge. Thanks to
  MPEG standard that leaves an intentional
  vagueness to allow for competing implementations
  of encoder part, a simplified psycho-acoustic
  model is utilized to reduce the computational
  requirements and memory consumption while
  preserving the coding quality.

21
Developing a MPEG Audio Layer III Codec on
TMS320C549An overview of the MPEG audio layer
III encoder algorithms is described in a block
diagram below
22
Developing a MPEG Audio Layer III Codec on
TMS320C549An overview of the MPEG audio layer
III decoder algorithms is described in a block
diagram below
23
Developing a MPEG Audio Layer III Codec on
TMS320C549

• The MPEG audio reference source code from ISO was
  written in C language. However since we want to
  implement a real-time system, we chose assemble
  language to develop our system.
• Spectrum Digital, Inc. provides the TMS320C54X
  Evaluation Module (EVM) that comes with the
  fixed-point TMS320C549 DSP on stand-alone card
  that lets evaluators examine certain
  characteristics of the TMS320C54x digital signal
  processor (DSP) to determine if this DSP meets
  their application requirements. Furthermore, the
  module is an excellent platform to develop and
  run software on the C54x family of processors.

24
Developing a MPEG Audio Layer III Codec on
TMS320C549

• The reference C language source code for MPEG
  audio layer III was implemented with regards to
  understanding the algorithms and total system of
  the standard, and not for speed. Moreover the
  computational type of C language is
  floating-point. To reduce the code size,
  complexity and also memory consumption, we
  decided to limit our encoder to16 bit and only
  support dual channel mode. We will discuss some
  of the major optimizing we have done in the
  sections below.

25
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Huffman Decoding We get the necessary
  information about huffman decoding from
  side-information part of every granule. Then we
  can decode the Huffman bit stream in each region
  according to the Huffman tree which will expedite
  the decoding rate compared with traditional table
  search. To avoid the too large Huffman table,
  linbits skill that offers extra bits to the large
  value in normal size table is used here. Because
  of the same properties of two kind of tables
  which have different linbits in most fields, we
  can save many table memory. In order to look up
  the tables quickly, there is a new table that
  contains the head addresses of every table. The
  decoder look up this table to determine which
  Huffman table is selected according the side
  information.

26
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Quantization Quantization part costs massive
  computation requirement that will impede the
  real-time realization of encoding, so we do some
  simplification on it linking with the change of
  psycho-acoustic model. Three layer nesting loops
  will be reduced to two layer loops, that decrease
  the computation explicitly. Surely, the quality
  of output compressed data degrades objectively,
  however the subjective listening result is fairly
  good. The significant meaning of this proposal
  makes the real-time MP3 encoding available to
  80MIPs DSP chip.

27
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Dequantization Owing to 16 bit fixed-point DSP
  chip used, the precision should be considered.
  Especially in dequantization part, because there
  is exponent function in this module. The exponent
  that has power 2 contains global-gain,
  subblock-gain, scalefactor and so on. We find the
  final exponent has the characteristics that 4
  times exponent is always integer, so the solution
  of this exponent function can include two parts.
  The first one is the integer part of the
  exponent. Because of power 2, we can use shift
  instruction to realize it. The other one contains
  six index (2 0.25 , 2 0.5 , 2 0.75 , 2 -0.25 , 2
  -0.5 and 2 -0.75 ) which will be multiplied to
  integer part if decimal fraction is not zero.
  This mixed method avoids the large exponent
  function table.

28
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Subband Synthesis Filters The subband synthesis
  filter bankis an inverse transform of subband
  analysis filter bank. It consists of an
  initialization section, the IDCT section, and a
  windowing section. The subband filter receives 32
  sub-bands of one channel and returns 32
  consecutive audio samples. In each iteration of
  the sub-band filter loop, the IDCT receives 32
  samples and returns 64 samples. The output
  samples is written to the FiFo buffer of 1024
  elements for the windowing operation. Thus, the
  windowing operation is done on the last 64
  elements together with the results from the
  previous IDCTs. A new vector is built from these
  elements and is multiplied with the windowing
  coefficients in the windowing section. The
  elements are then formed as a pulse code
  modulation (PCM)output. In the next iteration,
  all the elements are shifted down 64 places in
  preparation for the next IDCT output.

29
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Subband Synthesis Filters subband synthesis
  filter demands a large amount of computation and
  we should concentrate our efforts on this block
  to be able to build a real-time encoder. A fast
  transformations was developed for DCT32-gt32 and
  then corrected to the MPEG standard. As described
  above, MPEG audio decoding uses a IDCT 32-gt32.
  The direct implementation of 32x32 IDCT requires
  1024 multiplication and 992 additions. Therefore,
  fast DCT algorithms are used. Conventionally, FFT
  is utilized to reduce the computational
  requirements. However, we find that Lees FDCT is
  faster than FFT with trading away a little of the
  quantization precision. After careful
  consideration, we select FDCT scheme to implement
  this transform.

30
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Subband Analysis Filters The subband analysis
  filter consists of an initialization section, a
  windowing section and, the DCT section. This
  block also requires some optimization because it
  consumes a long time process too. The MPEG audio
  standard uses a DCT64-gt32 to divide the time
  samples into 32 subbands. Due to the symmetry of
  transform coefficients, we can convert input 64
  samples into 32 samples. Then the DCT64-gt32 can
  be transformed into a normal DCT32-gt32. Utilizing
  the similar procedure described in the part of
  subband synthesis filters, we can implement the
  DCT transform drastically reducing the
  computational requirements and the memory
  consumption.

31
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Psycho-acoustic Model Psycho-acoustic model
  calculates the Signal to Mask Ratio (SMR) for
  each subbands of input samples. The bit and noise
  allocation operation uses this SMR to determine
  the quantizational method for each subband. Due
  to the high computational requirement and memory
  consumption in psycho-acoustic model, it is a
  huge obstacle for real-time implementation.
  Furthermore, psycho-acoustic model analysis
  contains many floating-point calculations that
  are very difficult to implement on fixed-pointed
  DSP. Therefore, in our scheme of the MPEG
  encoder, we do not use the psycho-acoustic model
  analysis. By ignoring that, we do not calculate
  any block type and the SMR ratio is never set.
  Not using the SMR will also reduce much
  computation time in iterative loop of bit and
  noise allocation. This saves a lot of
  computations and memory, but trades off some of
  the high quality in theory.

32
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Decoder Results Below are some of the decoding
  results and technologic specification.
• 1. In order to test the robust of our decoding
  part which should be able to play any bitstreams
  supported by MPEG audio standard, we downloaded a
  Layer III test bitstream package from Fraunhofer
  IIS which contains several 10 strange
  bitstreams. All of them can be decoded
  successfully by our decoder.
• 2. In order to test the decoding quality of our
  decoder, we use a testing bitstream also
  downloaded from Fraunhofer IIS compl.bit
  (10Hz-10kHz/-20dB sine sweep, mono, 48 kHz). Our
  decoding wave file is compared with the ideal
  wave file and obtained the score of 93.34dB while
  the required score is only up to 77dB. At the
  same time, we also use floating-point C decoding
  program offered by MPEG organization to decode
  this test bitstream and the score is 93.35dB,
  almost same as our result.

33
Developing a MPEG Audio Layer III Codec on
TMS320C549

• 3. Memory consumption
• 1). ROM Program 4.75k Word, Parameter Table
  8.50k Word Total 13.25k Word
• 2). RAM Variables 11.00k Word, Total 11.00k
  Word
• 4. Computational complexity Because the
  computing complexity is partly determined by the
  bitrate, sampling frequency, stereo mode of Mp3
  source bitstream and the computing precision. We
  count the computing complexity using a common
  testing bitstream. Bitstream funky.mp3 with
  bitrate 96kbps and sampling frequency 44100hz.
  Its stereo mode is Joint Stereo. Under the 32bit
  computing precision, the computing complexity is
  about 52.3MIPS.

34
Developing a MPEG Audio Layer III Codec on
TMS320C549
35
Developing a MPEG Audio Layer III Codec on
TMS320C549
36
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Encoder Result
• 1. Memory consumption
• 1). ROM Program 4.10k Word, Parameter Table
  16.20k Word Total 20.30k Word
• 2). RAM Variables 10.60k Word, Total 10.60k
  Word
• 2. Computational complexity Because the
  computing complexity is partly determined by the
  bitrate, sampling frequency, stereo mode of Mp3
  source bitstream and the computing precision. We
  count the computing complexity using a bitstream
  with a bitrate of 128kbps and sampling frequency
  of 44100hz. Coding channel mode is dual channel.
  Under the 16 bit computing precision, the
  computing complexity is about 61.7MIPS.

37
Developing a MPEG Audio Layer III Codec on
TMS320C549

• Encoder Result
• 3. Encoder Quality Due to lack of software
  testing coding quality, we tried to code several
  wave files using our encoder system. Through our
  informal listening test, we can not discriminate
  the difference between original wave and coded
  MPEG audio bitstream.

38

• Thats All.
• Thank Everyone!