Inside IP-Voice Compression

Suppliers of IP-voice compression are coming up with new
ways to reduce latency that will bode well for the cable industry's ambitions in packet
telephony.

Any delays in the flow of Internet-protocol packets caused
by compression and decompression -- which is necessary to convert signals to and from
native formats -- can cause calls to become garbled because of latency.

And the effect -- which isn't unlike walkie-talkie-style
communication -- is worsened when combined with traffic-routing and distance-induced
delays that are intrinsic to data networking.

As a result, many people besides cable engineers are
looking for solutions that shave a few milliseconds off at the points of signal
processing, often in conjunction with moving to lower compression ratios.

Kalpana Sheth, director of marketing communications with
Lucent Technologies' Telemedia group, said the IP-telephony market is currently demanding
solutions operating at 7.3 kilobits per second and 9.6 kbps, "which is where we think
conferencing applications are going."

Neither of these rates falls within the codec
(coder/decoder) rates embraced by the International Telecommunications Union for
standardized voice-over-IP applications under its H.323 regime. There, the most common
audio-compression rates are 5.3 kbps and 6.3 kbps (G.723 class).

But that doesn't matter, noted Dror Nahumi, president of
MiBridge Inc. and vice president of emerging technology for MiBridge's parent,
IP-telephony carrier I-Link Inc..

MiBridge is shipping software tool kits for building
applications at these higher rates, Nahumi said. "At these rates," he noted,
"the packets are less sensitive to noise, and that's really important, especially for
conferencing."

A new focus on improved quality is good news for the
PacketCable group, which is tasked with coming up with voice-over-IP specifications for
the cable industry. Cable can afford to give up some bandwidth in the interest of improved
quality, PacketCable executives said last week.

That's especially true if the use of less compression, or
even no compression, improves both the quality of the sound and the latency performance,
noted Frank Christofferson, project director for software systems at Cable Television
Laboratories Inc.

"We can't afford latency in excess of 20 milliseconds
in the compression and related functions," Christofferson said. "Anything higher
pushes against our end-to-end goal of achieving less than 200-millisecond latency for
toll-quality IP voice."

These parameters rule out the use of G.723.1-level
compression, which adds delays on the order of 67.5 ms. Another format, G.729, operates at
8 kbps. It is closer to the target, imposing latency of only 25 ms, Christofferson said.

While jumping to G.728 at 16 kbps with only 1.25 ms latency
would eliminate the problem, that's a huge price to pay in added bandwidth consumption
when G.729 is so close, he added.

Thus, for cable, squeezing that last few milliseconds out
of the latency at G.729 could mean savings in the tens of millions of dollars when it
comes to infrastructure improvements that will be needed to support high levels of
penetration for future IP-voice services.

Latency improvements are one of the benefits to be realized
from Inter-Tel Inc.'s recent purchase of Telecom Multimedia Systems Inc., said Jeff Ford,
chief technical officer at Inter-Tel, a full-service provider of digital-telephone systems
and software, including IP-telephony systems.

"We already had an IP gateway, Vocal'Net, which is
doing well for us in the carrier market," Ford said. "But it didn't fit the
right price points for payback in the business market over private wide-area
networks."

TMSI had already developed ways to integrate compression,
echo cancellation, jitter management and IP-call setup onto DSP (digital-signal
processing) boards, Ford said. Because of that, Inter-Tel was able to produce stand-alone
boxes the size of small routers that attach to PBXs (private-branch exchanges), avoiding
the costs of using Intel Corp.-class computer platforms to accomplish these tasks, he
added.

Moreover, he said, the tight integration on the fast DSPs
supplied by Motorola Inc. helps to cut latency to under 20 ms from the point of call input
to the call exit.

The resulting products are positioned to meet the surging
demand for IP-voice solutions in the corporate domain, allowing businesses with as few as
40 phones to benefit from the voice-over-data capabilities that IP telephony provides.
"A corporation serving remote offices with a 256-kbps data link can supply four
channels of voice riding on the data channel nearly for free," Ford said.

The next step for Inter-Tel involves producing
standards-compliant systems. "Our H.323 systems will benefit from the same low-cost,
high-quality attributes that we've gained from the TMSI technology in our enterprise
products," Ford said.

One of the techniques that Inter-Tel is using for reducing
latency, along with tight integration, involves a reduction of the frame size for the
IP-voice packets, which run at 30-ms spacing in G.723, for example. The Inter-Tel
"Interprise" system employs 10-ms framing using the TMSI proprietary compression
algorithms, which happens to be the frame size used in G.729 algorithms, as well.

Reduced frame size is definitely a plus for G.729, leading
to end-to-end latencies in the PBX-to-LAN (local-area network) enterprise domain of under
100 ms in a fully integrated system, Nahumi noted. "This is one of the reasons why
we're seeing more demand for the higher-bit-rate solutions," he said.

MiBridge has partnered with Brooktrout Technology Inc. in
developing Brooktrout's TR2001-series fax and voice platform as a resource board for
H.323-compliant applications in the enterprise and carrier markets.

Capitalizing on the performance gains in DSP technology,
the integrated board can support up to 60 channels of IP-voice and fax processing for
gateway systems that will eventually find their way into higher-capacity public-access
links such as xDSL (digital-subscriber-line) facilities, officials said.

Another area of development for IP-voice systems is the use
of wideband sampling rates: Rather than reducing frame sizes, vendors increase the
data-sampling rate of the host processor. This is a key focus of work under way at Voxware
Inc., which made its name by breaking through to ever-lower bit-rate coders for the
IP-voice market.

"There's a big need for high-quality,
wideband-telephony applications that are impervious to packet loss in the private Intranet
market," said Peter Gantchev, market manager for voice-over-IP products at Voxware.
Gantchev said carriers like Qwest Communications International Inc., which are loaded up
with bandwidth, are interested in distinguishing their products with high-quality sound.

Voxware's wideband "Excelsior" system, which is
slated for product release next year, delivers "near-transparent" sound quality
using 8-to-1 level, 16-bit compression and 32-kilohertz sampling. Current narrowband
efforts use much higher compression ratios, with 8-bit coding and 8-KHz sampling, Gantchev
noted. The new approach leapfrogs the current ITU benchmark for wideband applications,
which uses 16-KHz sampling.

"The ITU is going back and forth on wideband,"
Gantchev said. "I think that they'll go to 16 KHz first, but that they will be
quickly pulled by market developments to 32 KHz."

Wideband sampling with lower compression ratios assumes
that much more bandwidth is available, of course, than what is necessary for narrowband
voice-over-IP applications. But it can be remarkably bandwidth-efficient, allowing users
to benefit from the advantages of IP over data while accomplishing the pristine sound
quality that corporate users are used to with circuit-switched PBX and conferencing
systems.

In Voxware's case, the new 32-KHz wideband product will
packetize 64-kbps voice signals from the circuit-switched domain, while reducing bandwidth
consumption by a factor of eight for transport over the data network, without sacrificing
quality, Gantchev said.

"Ideally, what will happen is that all users on the
LAN and in branch offices connected over the high-speed Intranet will be equipped with LAN
phones so that the IP-voice connections never touch any 8-KHz circuits," he added.

The willingness of packet-voice-system customers to
experiment outside of the standards framework is good news for cable, which is seeking
both to use standardized gear and to foster new standards for cable-specific applications.
Demand for nonstandard solutions generates products that might be useful for cable, which
wouldn't be out there otherwise.

"There are a lot of applications that don't require
interconnection with another network, where people are looking for optimized solutions
without having to worry about paying the royalty costs," noted Bruce Gellman, vice
president of AudioCodes Ltd., which supplies chips and software for compression
applications.

AudioCodes has taken a novel approach to the
noise-versus-bandwidth conundrum by introducing what it calls the "NetCoder" --
a proprietary coder that dynamically adjusts the bit rate to anywhere from 4.8 kbps to 9.6
kbps, depending on line conditions. At press time, the company was about to announce
shipments of multiple-coder technology to IP-systems supplier Hypercom Corp. for use in
that company's new enterprise-networking software.

The AudioCodes compression chip combines toll-quality
low-bit-rate voice compression, G3 fax relay, silence suppression, echo cancellation and
all voiceband-processing functions necessary for implementing voice-over-packet products,
Gellman said.

Along with NetCoder, the chip supports four independent
full-duplex voice/fax channels with user-selectable G.729A, G.723.1, G.727/726 and G.711
codecs.

Application integration at this level, combined with high
channel density, is crucial to the goal of reaching carrier-class scales in gateways and
other voice-over-IP components, Gellman noted. "High density is the name of the game
to bring IP voice to the carrier marketplace," he said.