Like an emergency fire crew with access to hundreds of gallons of water but only a garden hose to deliver them, today's network designers find themselves with terabits of backbone capacity but only a kilobit-per-second pipeline to their customers. Replacing the local copper loop with fiber will be expensive and take decades, so until then, the main challenge is to optimize the pipeline in the local loop so that it can deliver significantly more available bandwidth.
Estimates suggest that backbone bandwidth, including packet-switched and circuit-switched networks, is doubling every nine months. Bandwidth to the customer is growing much more slowly, although recent developments in digital subscriber line (DSL) technology have increased the achievable data rates over copper. The new HDSL2 standard, for example, will allow T1 or E1 operation over a single twisted pair, something that previously took two or three twisted pairs.
Although additional gains may be achieved with better modulation techniques or by canceling crosstalk from other high-speed services, significant additional gains are probably not possible, so other avenues must be explored. Any technique that exploits signal processing, such as speech coding, is an especially attractive option because it can be done affordably in silicon and is a one-time-only capital-equipment cost.
A promising technique is to mix voice and data over a DSL line, especially an HDSL business access line. Known as voice over DSL, or VoDSL, this method uses a low-delay speech coder for the voice plus a communications protocol, such as asynchronous transfer mode (ATM) or Internet protocol, to intermix voice and data packets. Significant additional capacity can be achieved by using silence suppression on the voice lines.
The customers most likely to deploy VoDSL are small to midsize businesses. Because large companies usually require many voice lines and very-high-speed Internet access, they are more likely to choose dedicated lines rather than lines that combine both types of access. Small to midsize companies, on the other hand, probably need fewer than 100 voice lines, along with high-speed Internet access. VoDSL allows a competitive local exchange carrier to supply those companies with many different communications services over a single local loop.
VoDSL appeals to customers who like the idea that the bandwidth doesn't remain idle simply because no one is using the phone. The bandwidth of T1 voice-access lines is wasted most of the time, because even active channels are communicating silence when no one is speaking.
In a typical VoDSL application, a piece of equipment that contains some number of speech coders at the customer premises communicates with a DSL access multiplexer (DSLAM) in the central office. ATM is typically the underlying protocol. The ATM cells are routed through the ATM network to a voice gateway, where the voice lines are converted to time-division multiplexing (TDM) circuits and connected to a standard 5ESS or DMS-100 telephony switch. Data cells are routed to the Internet at the gateway, or the Internet connection could occur at the DSLAM. In either case, ATM adaptation layer 2 (AAL2) cells carry voice and AAL5 cells carry data.
The voice circuits can serve the employee's telephones directly, the way a Centrex system would, or they can act as trunk lines to connect the company's PBX to the public-switched telephone network.
When voice is communicated continuously it is easy to calculate how much bandwidth is required to support a certain number of voice calls: Multiply the bandwidth required for one voice circuit by the number of simultaneous calls. For example, the PSTN uses 64 kbits/second for each phone call. Two simultaneous calls require 128 kbits/s, four simultaneous calls require 256 kbits/s, and so on.
When silence suppression is deployed, bandwidth is used for voice only when someone is speaking. Picture an airport ticket counter: Suppose there was only one ticket agent but two lines, one for first class and one for tourist class. When there is no one in the first-class line, the ticket agent serves the tourist-class line. But as soon as people arrive in the first-class line, the agent serves them.
Think of the voice samples as the first-class passengers, the Internet data as the tourist-class passengers and the ticket agent as the HDSL line communicating cells. The network gives priority to the voice cells because they are sensitive to delay. But it is important to understand the average arrival rate of passengers into the first-class line; that is, how long first-class passengers-or in our example, voice samples- will have to wait, and how many tourist-class passengers or Internet data transmissions the agent can facilitate on average.
The first issue is fairly straightforward. In 1969, a study of telephone network speech patterns showed that in a typical conversation, 352-millisecond talk spurts were usually followed by 650 ms of silence. After exponentially distributing the values over the long term, it was determined that a person speaks about 35 percent during a phone call. With that knowledge, it's possible to take the bit rate of the coder and compute the number of AAL2 cells that would be generated by a certain number of concurrent speakers per unit of time. That is equivalent to the average rate of first-class passengers entering the ticket line.
Going back to our example, once the average arrival rate of first-class passengers is determined, it's possible to calculate the average percentage of time the ticket agent will spend serving first-class passengers. Armed with that figure, you can then calculate the average number of tourist-class passengers (AAL5 data cells) served, because the remaining time will be used to serve tourist-class passengers.
Calculating how long a first-class passenger will have to wait without resorting to queuing theory mathematics would be difficult. Given system conditions, however, one can at least calculate the probability of a first-class passenger having to wait more than a certain period of time. Most companies developing VoDSL systems believe that the system should not add any more than 20 ms of one-way delay into the call, so there should be a very low probability of a speech sample being in the queue more than 10 ms.
It's possible to use a number of speech coders, but G.711, which operates at 64 kbits/s, is the standard. Although G.711 has many advantages, it is possible to communicate speech with high quality at lower bit rates. G.726 can reduce the bit rate to 32 kbits/s and still provide very good voice quality. G.728 further reduces the bit rate, to 16 kbits/s, and provides speech quality as good as, or better than, G.726 while still providing very low delay.
Using G.728, Conexant simulated the VoDSL system using MathCAD. The results showed that a different number of voice lines can be established depending upon how much bandwidth is reserved for data. For example, in the T1 (1,544-kbits/s) section, the simulation showed that the line could support 99 simultaneous voice circuits and still provide an average of 729 kbits/s for Internet access. This is a significant improvement over 24 voice circuits and no data.
When speech coders are used for voice, however, several problems arise. A network uses many tones: dual, multifrequency for dialing and call-progress such as "busy" and "ringing," including foreign call-progress tones for international calls. If the voice lines are used as a "trunk" to a PBX, multifrequency tones may be put on the line. If the lines are used to provision telephony service directly, custom local area signaling services (Class) such as Caller-ID must be passed through. All of these tones and signals must be detected and passed by the speech coders. G.728 does a better job than lower bit-rate coders, but special techniques are required for reliable operation.
Because these derived voice circuits behave like ordinary lines, customers may use one or more lines for fax machine support or for plain old telephone service modem operation. Unless these uses are detected, the fax or modem will probably not operate correctly. One solution is to switch the line to G.711 (64-kbits/s) operation. However, this redirects bandwidth from the voice circuits, significantly lowering the number of lines that can be supported.
Another possible solution is to use "demod-remod" techniques, in which the analog signal is detected, and the digital data encoded in it is extracted. Only the digital data is then communicated over the HDSL circuit.
The copper local loop remains a network bottleneck, and while innovative techniques have increased available bandwidth, huge barriers to additional increases exist. Yet, even though there is a limit to raw bandwidth in the local loop, other techniques can expand services. The combination of voice and data traffic over the same HDSL line, coupled with the use of G.728 and silence suppression, will increase the number of voice circuits that can be provided.
See related chart