18-796 Multimedia Communicaions:
Coding, Systems and Networking

Final Project Report

Group members:     Ching-Kai Huang         Wing Ho Leung

Implementation of a speech codec based on coding of speech at 8 kbit/s
using conjugate-structure algebraic-code-excited linear-prediction (CS-ACELP)
[ITU-T Recommendation G.729A]

    Our goals for the final project is to implement the encoder and decoder following the G.729A recommendation except for some quantization modules and bit stream compliance.

    We took .wav files that is sampled at 8000 Hz using 16-bit linear PCM. The encoding process is done every 10ms frame or 80 samples. For the preprocessing stage, the samples are high passed with cut-off frequency of 140 Hz and scaled down by 2. A total of 240 samples are buffer for windowing and autocorrelation computation. The autocorrelation coefficients are used to calculate the LP filter coefficients using the Levinson-Durbin algorithm. The LP filter coefficients are converted to Line Spectral Pair (LSP) coefficients.  LSP coefficients are converted back to the LP filter coefficients, which is just the reverse process of the conversion from LP to LSP. This module is exactly what the decoder will need in order to convert the LSP coefficients to LP coefficients. We decided not to implement the LSF quantization module because we did not have the codebook information when we designed our system.
The open-loop pitch delay is calculated first for each frame.  Then the closed-loop pitch delays for each subframe are estimated around the open-loop pitch delay using the correlation function calculated from the pre-processed signal. The adaptive code vector is then obtained by interpolating the excitation signal around the closed-loop pitch delay. Next, the fixed code vector which consists of 4 pulses are obtained by searching the codebook and selecting the one whch maximizes the correlation term normalized by it's energy. The fixed code vector is passed to a pitch pre-filter if the closed-loop pitch delay is small. The adaptive and fixed codebook gains are calculated and the excitation signal and the states of the weighting filter are updated.

    The modules of the decoder are similar to the inverse of the encoder, and even simpler. The idea is that based on the encoded parameters, we reconstruct both the adaptive and fixed code vectors, then scale them with the adaptive and fixed codebook gains which are also parameters from the encoder. Then we calculate the excitation signal by adding the scaled adaptive and scaled fixed code vector. On the other hand, we also find the LP coefficients from the LSF parameters. We use Linear Prediction to reconstruct the speech signal using the excitation signal and the LP coefficients. Then the signal is passed to the post-processing stage to improve the speech quality.

    The following table shows our parameters to be transmitted for each frame:
 

    Without coding, 16 bits would be required to represent each sample and therefore 1 frame which corresponds to 80 samples requires 16x80 = 1280 bits.  From the above table, it can be seen that 496 bits are used to code 1 frame. This gives a compression ratio of 2.58. We haven't made use of the VQ tables for coding line spectral frequencies as L0, L1, L2 and L3; and haven't coded the codebook gains as GA1, GB1, GA2, and GB2. Otherwise each frame will only require 80 bits and the compression ratio would become 16.

    We have modified our MATLAB graphical user interface built before the midterm. This user interface is enhanced to be capable of displaying different signals of different frames. The signals generated from the encoder to be displayed include the original speech signal, the preprocessed signals, the LP residual signal, the adaptive and fixed code vectors, the excitation signal and Line Spectral Frequencies. The signals generated from the decoder to be displayed include the reconstructed LP residual signal, the reconstructed adaptive and fixed code vectors, the reconstructed excitation signal, the signal before postprocessing and the reconstructed speech signal. Two different graphs can be viewed at a time in order to facilitate comparison. In addition, we can listen to the original and reconstructed .wav files using our MATLAB user interface in order to evaluate the quality of the reconstructed speech. The following figure shows a snapshot of our MATLAB graphical user interface:

    The source codes for this project are:
           encoder.zip     - program source code to encode a WAV file
           decoder.zip     - program source code to decode back to a WAV file
           get_signal.zip   - program source code to get intermediate signals for both the encoder and the decoder to be used by the MATLAB demo
           demo.zip         - demo directory which contains the EXE files for the above programs, the MATLAB demo files, and some test WAV files.
                                      also contains a readme.txt file to help explaining how to use the demo programs.