^8 8 - 25 75 5
DEVELOPMENT OP AN 8000 BPS VOICE CODEC FOR AvSat
JOSEPH F. CLARK, AvSat Program, Aeronautical Radio, Inc., United States
AERONAUTICAL RADIO, INC.
2551 Riva Road
Annapolis, Maryland 21401
United States
ABSTRACT
Air -mobile speech communication applications
share robustness and noise immunity requirements with
other mobile applications. The quality requirements
are stringent, especially in the cockpit where air
safety is involved. Based on these considerations, a
decision was made to test an intermediate data rate
such as 8.0 and 9.6 kb/s as proven technologies.
A number of vocoders and codec technologies were
investigated at rates ranging from 2.4 kb/s up to and
including 9.6 kb/s. The "proven" vocoders operating
at 2.4 and 4.8 kb/s lacked the noise immunity or the
robustness to operate reliably in a cabin noise envi-
ronment. One very attractive alternative approach was
Spectrally Encoded Residual Excited LPC (SE-RELP)
which is used in a multi-rate voice processor (MRP)
developed at the Naval Research Lab (NRL). The MRP
uses SE-RELP at rates of 9.6 and 16 kb/s. The 9.6
kb/s rate can be lowered to 8.0 kb/s without loss of
information by modifying the frame. An 8.0 kb/s
vocoder was developed using SE-RELP as a demonstrator
and testbed. This demonstrator is implemented in
real-time using two Compaq II portable computers, each
equipped with an ARIEL DSP-16 Data Acquisition Pro-
cessor .
INTRODUCTION
The Airlines Electronic Engineering Committee (AEEC) has undertaken
development of a "characteristic" for satellite communications equipment
— in effect a guideline for equipment and airframe manufacturers,
aviation users, and service providers. Designated Project 741, it is
being developed in parallel with a wide variety of intense activities in
the mobile satellite communications area.
The primary schedule constraint for the AEEC has been introduction
of a new generation of aircraft to be delivered "satcom ready" starting
in 1989. As a consequence, the AEEC formulated a three-phase schedule
for Project 741. By late 1986, the Satellite Subcommittee of the AEEC
formed a working group to recommend a voice coding standard. Although
there has been a hiatus in the activity of that working group, its
521
efforts both in fulfilling its mission and in assuring that all aspects
of the voice communication system be considered together are continuing.
This paper reports on the specific effort to develop a voice coding
standard using the transmission rate of 8.0 kilobits per second (kb/s)
and focuses on the similar performance, as shown in testing, of the 8.0
kb/s and 9.6 kb/s codecs available today.
ADAPTING TO THE USER COMMUNITY
The aviation conmunity is loosely divided between "general avia-
tion” and commercial passenger aviation. The aircraft used vary in size
from the very large to the very small. The length of flights varies
from less than one hour to more than eight hours. Although data commu-
nications increasingly is the mechanism for reporting aircraft status
and for routine operational exchanges, one common denominator is a
requirement for cockpit voice — a communication path for the flight
crew to talk with ground crews.
While the primary volume application of voice appears to be passen-
ger telephony, the cockpit voice quality requirements may be the primary
determinant of the overall requirements because of the ties to safety.
While only experimental equipment has been installed to date, it is
clear that aviation will use both high-gain and low-gain antennas. The
service providers will be using a mix of global beams and spot beams
from the satellite. As a result, three classes of service can be pro-
vided: low-rate data; medium— rate data and low— rate voice; and the
nominal, which is medium-rate data and "near-toll-quality" voice.
We suggest that a single algorithm for both low— rate and medium-
rate voice would be the preferred solution. This approach was developed
by the Naval Research Laboratory (NRL) 1 and is in operation today. The
method for providing voice communication through a low-rate data service
is to buffer short messages coded at the nominal voice rate, and trans-
mit them as extended messages at the actual channel data rate.
PERFORMANCE REQUIREMENTS
So far, no one has defined well enough the measurable parameters
for repeatable measurements of "near toll quality." However, the
achievable performance of codecs is well known. Figure 1 depicts the
achievable performance of a wide variety of codecs as determined in
tests. But in order to specify a required performance, it is necessary
to determine minimum acceptable performance in terms of parameters that
can be measured in a consistent way. Unfortunately, voice quality is by
its very nature subjective. It can only be quantified as a statistic
derived from inherently variable data.
The performance requirements endorsed by the AEEC Voice Working
Group^ are simple:
Subjective evaluation measured by Mean Opinion Score - midway
between Fair and Good (3.25 on a scale of 1-5).
Intelligibility measured by Rhyme Test - 87 or better.
- Conversational evaluation - 75 or better.
Delay measured with codecs back to back - no more than
65 milliseconds.
522
Performance measured with a channel bit error rate (BER) of 1 in
1000 bits.
In addition, the group recognized that the operational environment
would include low-rate data transmissions, rather severe outages and
interfaces with established terrestrial telephone networks. It was
decided that data should be routed around the codecs. Not only does the
requirement for coding data as well as voice put significant demands on
the coding algorithm, it implies significant additional acceptance test-
ing.
The new CCITT coding standard for digital transmission of wideband
audio signals, G.722, provides a similar bypass mechanism for user data.
The separation of voice and data allows a better system design 3 .
APPROACHES
A CCITT standard ADPCM would simplify the task, but there is more
complexity in the transmission channel. Limitations of the medium, and
power and bandwidth constraints drive the choice of rates to the lowest
possible. On the other hand, the realizable quality at a given rate
drives the choice upwards. The result then, is a compromise on an
intermediate rate that approximates the desired performance without
devouring all available capacity.
523
The rate issue has resolved to a choice between 8.0 kb/s and 9.6
kb/s. The 8.0 kb/s rate is based on compatibility with expected
subchannelling of the ISDN basic 64 kb/s communication rate.
Intuitively, there should not be much difference in performance at these
rates relative to, say, performance at 16 kb/s or at 4.8 kb/s. This
intuition is supported by the evaluations received so far. (See Figures
2 and 3.)
Furthermore, performance differences seem to be related more to
implementation than to algorithm. This result is quite reasonable since
nost of the algorithms are variations on the idea of using residual
excitation with a linear prediction filter.
The conclusion then is that the AEEC's voice working group could
not recommend a specific algorithm, or even a data rate. Only
implementations of voice systems can be evaluated and recommended.
From an overall systems point of view, there is a significant
difference between these rates. Choosing 9.6 kb/s without gaining a
significant performance advantage is wasteful. In the best of proposed
implementations, 9.6 kb/s uses $1.20 worth of resources for every $1.00
used by 8.0 kb/s. For SC PC voice channels as proposed in Project Paper
741, the 9.6 kb/s implementation requires almost twice as much bandwidth
as the 8.0 kb/s implementation. (See Table 1.)
*
Table 1. Voice Channel Frame Parameters
Voice
Rate
(bits)
FEC
Rate
Channel
Rate
(bit/s)
Channel
Spacing
KHz
Bits in Frame
Voice/25
(v)
Data/96
(SU)
(n)
Dummy
(d2)
9600**
8000**
0.50
0.75
21000
12600
17.5
10.0
192
160
3
6
84
8
♦Extracted from Project Paper 741, Part 2.
♦♦Preferred Voice Channels
Hie AEEC voice working group with the assistance of the Boeing Air-
plane Corporation developed a standard test tape. This tape was
distributed to approximately twenty— five organizations which had expres-
sed an interest either in developing codecs for the aviation industry or
in evaluating voice codecs. The tapes were fed in and out of candidate
codecs, and the resulting tapes analyzed in listening tests conducted by
Dynastat.
CNE 8.0 KB/S SOLUTION
There were two major concerns in dealing with established manufac-
turers of voice codecs. First, a satisfactory licensing agreement for
the algorithm and implementation must be available to all avionics manu-
facturers. Second, manufacturers might continue to focus on products
designed only for their terrestrial markets, i.e., codecs designed
524
Fig. 2. Intelligibility: Cockpit
Noise Environment
Fig. 3. Probability of User Acceptance Based
Based on Cockpit Noise Environment
for 2.4 kb/s, 4.8 kb/s, and 9.6 kb/s even though 8.0 kb/s is a
demonstrably better choice for the emerging aviation industry character
istic .
In view of these concerns, ARINC contracted with Techno-Sciences,
Incorporated (TSI) of Greenbelt, Maryland, to develop an 8.0 kb/s voice
codec from information published in the public domain. The premise was
if a credible voice codec could be developed by one or two key people
working with limited time and budget, then avionics manufacturers could
derive non-proprietary sources for voice codecs, and on a realistic
schedule with confidence.
525
TSI, with ARINC assistance, developed a pair of 8 kb/s voice codecs
operating in real time. The codecs are connected through the serial
(RS232) ports of two Compaq Portable II computers. Speech is sampled
8000 times a second and blocked into groups of 180 samples. Ten
reflection and filter coefficients are calculated using autocorrelation.
The quantized reflection coefficients are used to form the filter which
then is used to generate the prediction residual.
A 96-component spectral representation of the prediction residual
is computed using the Winograd algorithm. These components are
quantized and coded into a 180-bit block with the reflection
coefficients. A new block is generated and transmitted 44.4444 times
per second. The key factors in the codec performance are the reflection
coefficient calculation and the number of transmitted spectral
components. The current codec transmits 126 spectral components, the
same number transmitted in the NRL 9.6 kb/s rate codec.
After 6 months of development, a codec emerged that is credible but
admittedly is incomplete and needs improvement. Several areas of the
basic LPC algorithms have been tagged for further work in a second phase
of the project. In the second phase of the codec development, vector
coding of the spectral components will be introduced. There is good
confidence in the expected results.
CONCLUSION
The AEEC Voice working Group, with cooperation of the Boeing
Airplane Company, the Federal Aviation Administration, the Rome Air
Development Center and others, has developed a set of standards for
voice codecs in the air mobile environment, collected a representative
set of samples for a variety of oodecs using a uniform test tape, and
received an unbiased formal evaluation of the sample tapes. The results
of the evaluation indicate that at least 7.2 kb/s speech coding will be
required to reliably meet requirements in the aviation satellite
environment. These results (see Figures 2 and 3) also indicate that
performance at 8.0 kb/s is in the same range as performance at 9.6 kb/s
— making 8.0 kb/s the better overall choice for a power and bandwidth-
limited satellite environment to be globally interconnected with ISO-
conforming networks.
REFERENCES
1. Kang, G. S. and Fransen, L. J., Second Report of the Multirate Pro-
cessor for Digital Voice Communications, NRL Report 8614,
September 30, 1982.
2. AEEC Letter 87-014/SAT-33, p. 11, Aeronautical Radio, Inc., 2551
Riva Road, Annapolis, MD 21401, February 13, 1987.
3. Mermelstein, Paul, G.722, a New CCITT Coding Standard for Digital
Transmission of Wideband Audio Signals, IEEE Communications
Magazine, Vol. 26, No.l, January 1988.
4. AEEC 700 Quick Check #50, November 4, 1987.
5. Techno-Sciences, Inc., Report No. 870909, 8 kbps Speech Demo Pro-
ject, prepared for Aeronautical Radio, Inc., September 9,
1987.
526
