Speech coding

V 2.1, May 2, 2006

Dmitriy Shutin, dshutin@tugraz.at
Signal Processing and Speech Communication Laboratory, http://www.spsc.tugraz.at/
Institute of Communications and Wave Propagation, Graz University of Technology
Inffeldgasse 16c/II

Abstract
This is part of the laboratory is dedicated to Speech Coding. We will study princi-
ples of speech coding as well as speech coding quality assessment. The experiments
are performed in Matlab as well as in Praat.
Equipment:
• PC with a sound card installed
• Matlab and Praat software, a copy of laboratory files (these you can download from our
website
http://www.spsc.tugraz.at/courses/scl/download/coding.zip)
• Headsets and necessary cables.
Before you start...
1. You are expected to write a detailed report about your work. The report should be
prepared on the fly, as you proceed with the tasks and handed over at the end of the
laboratory.
2. Do not forget to answer the questions asked in experiments and those asked by the lab
assistant. Your answers should be provided in full, with all the necessary explanations
and, if required, plots and tables. Try to explain what you observe! If you can’t, do not
hesitate to ask your lab assistant!

Experiment 1
Linear models and quantization
In speech coding, the main goal is to represent the speech signal as compactly as possible,
preserving the audible quality. Speech modeling and proper quantization are the key elements
in achieving this goal.
This part of the laboratory is to be performed in pairs such that one of the group prepares
the speech material and another makes the measurement.
(1.1)LPC residual signal
Speech coding 2

1. Start Praat and load any speech file from speech coding/data directory into the Praat
workspace. The given files are in RAW format, i.e. without any header information. To
load the file follow Read → Read from special sound file →
Read from raw 16-bit Little Endian file... . Note: m38.raw and f68.raw files are in
German language while the others are in English.

2. For the raw sound file it is necessary to specify the sampling rate manually. Select
Modify - → Override Sampling frequency and set New Sampling rate to 8000Hz.

3. Press Formats & LPC → To LPC (autocorrelation) → Standards . This will create a
new LPC-object constructed from the selected signal.

4. Now, in the object browser select simultaneously the created LPC object and the cor-
responding sound signal. Press Filter (inverse) . The newly created object is the LPC
residual signal (in Speech Analysis laboratory we used it as an approximation to the
excitation signal).

5. Let the other person from the group listen to the created residual. Is it possible to
understand what was said? Compare it with the original sound.

6. Now, exchange your roles and repeat the above experiment.

7. Why the Can you tell why the LPC residual signal contains so much information?

(1.2)Quantization

1. In this task we continue to use the same experiment strategy, i.e. one person loads the
file and the other listens to the speech sample. Then the roles are exchanged.

2. Load any speech file from the speech coding/data directory into the Praat workspace.

3. Select the corresponding sound object and press Modify - → Formula... . In the created
dialog window specify the following expression: if (self)>0 then 1 else -1 fi.
This operation results in a 1-bit quantization of the original sequence.

4. Now, let the other person listen to the created sound. Is it possible to understand what
was said? Compare it with the original sound.

5. Repeat the experiment, but now exchange your roles.

6. What is your explanation of why the speech information is still audible?

Experiment 2
ADPCM
This experiment is dedicated to the ADPCM (Adaptive Differential Pulse Code Modulation)
coding scheme. ADPCM is an effective combination of the adaptive LPC modeling and quan-
tization.
(2.1)ADPCM coder
Speech coding 3

Figure 1: ADPCM speech coder

1. Start Matlab. Change the working directory to speech coding/adpcm where you will
find in data directory with input speech files and doadpcm.m script that will be pri-
marily used during the experiment.

2. Load doadpcm.m code into the Matlab editor. This script simply loads the specified
speech files and invokes the coder adpcm.m with the proper settings. There are two
important parameters that control the behavior of the script. The first one is the coder
bitrate, specified in the variable N (possible bitrates are 16,24,32, and 40 kbps), and
second parameter ExtendedOutput controls the number of output arguments in adpcm.m
function. In case ExtendedOutput is different from 0 the adpcm returns most of the
internal variables. Otherwise, only the coded index and decoded output signals are
returned.

3. Set ExtendedOutput to 1 and run the script. For every loaded file listen to the original
and the decoded signals. The corresponding speech samples are stored in the variables
x and y respectively. Using the built-in Matlab function sound play the decoded and
original samples. Note: in the Matlab implementation of sound function it is required
for the signal to be scaled between ±1.

4. For each played sample plot the corresponding original sequence x, difference signal d,
output of the coder dQ, and the difference between the original x and the reconstructed
signal y.

5. Perform the above analysis for bitrates N=16,24,32, and 40kbps.

Experiment 3
Objective coding quality criteria.
In this experiment we will study an objective quality measures to evaluate speech coding
algorithms. The difficulty in defining objective quality measure lies in non-stationarity of the
speech and strong speaker dependency.
(3.1)Segmental SNR
Speech coding 4

1. Using segsnr.m script from speech coding/quality directory estimate the segmental
SNR for the ADPCM coder from the previous experiment.

2. Run the doadpcm.m script and wait until the program enters the debug mode. The
original signal x and the decoded signal y are stored in the workspace.

3. Fill in the following table:

L msec / bitrate 16kbps 24kbps 32kbps 40kbps

5 msec
15 msec
30 msec
60 msec

Note: Before computing the SNR make sure that both the decoded signal y and the
original signal x have the same length and time aligned.

Experiment 4
Subjective coding quality criteria.
Subjective criteria, on the other hand, are measured by a group of listeners that give their
personal evaluations of the encoder quality. We will consider two of them – Modulated Noise
Reference Unit (MNRU) and Mean Opinion Score (MOS). We will also study some speech
coders that are used nowadays in digital communication systems.
(4.1)Modulated Noise Reference Unit

1. Just as in the previous experiment, run ADPCM encoding for the given speech signals.

2. For the original signal x generate the corresponding equivalent noise signal x eq using
mnru.m script from speech coding/quality directory. Set noise parameter Q between
10 and 40dB.

3. Compare perceptually the ADPCM decoded signal with the equivalent noisy signal x eq.
The value of Q that results in a similar speech quality is used as a quality measure.

4. Fill in the following table:

16kbps 24kbps 32kbps 40kbps

Q, dB

(4.2)Computing MOS

1. If necessary, start Matlab and load encode decode.m script from speech coding/quality/mos
directory. This script automatically encodes and decodes the specified speech signal with
the specified encoder. It requires a single parameter – the index N of the coder as shown
in Table 1.
Speech coding 5

N Coder Listener 1 Listener 2 MOS Total MOS

1 AMR 4.75 kbps
2 AMR 12.2 kbps
3 G722 6.6 kbps
4 G722 23.8 kbps
5 G729 6.4 kbps
6 G729 11.8 kbps

Table 1: MOS

2. Now, let one of the group choose a coder at random by specifying the corresponding
coder index. Run the script and, when prompted, enter the file name of a speech file
from speech coding/quality/mos/samples/ directory. When the script enters the
debug mode, play both the original and the decoded speech signals to you partner.
Note: the partner shouldn’t know what coder is used and what sound file is chosen. He
only has to judge to quality of the speech sample using the scale shown in Table 2. Once

Score Description
5 – excellent Imperceptible difference
4 – good Just perceptible but not annoying
3 – fair Perceptible and slightly annoying
2 – poor Annoying but not objectable
1 – unsatisfactory Very annoying and objectable

Table 2: MOS grade system

you finished, exchange you roles.

3. Repeat the above steps until the Table 1 is filed. The MOS column of the table is
computed as the mean value of the obtained scores for each coder.

4. The total MOS is computed as a mean score for the corresponding coders among the
groups, i.e. individual MOS scores are averaged together.