VoIP (Voice over IP)

Prior to Voice over Internet Protocol (VoIP), separate networks carried voice and data traffic: circuit-switched voice traffic and packet-switched data traffic. The two networks actually operated on the same types of wires, but the physical network infrastructure was optimized for the circuit-switched voice traffic because voice traffic existed first and accounted for the vast majority of the traffic.

Over the past 20 years, however, the volume of data traffic has increased exponentially. Studies suggest that data traffic began to exceed voice traffic some time between 2002 and 2003. Because bandwidth is a limited and therefore expensive resource, there is continuous pressure to use it more efficiently. One of the best ways to efficiently use bandwidth is to converge voice and data networks. Convergence also reduces training and operational costs because you have only one network to maintain. Because the primary network is now packet (data) based, it makes more sense to modify voice signals to traverse data networks than vice versa.

Converting analog voice signals to digital packets
All sounds (including speech) are analog waves of one or more frequencies. VoIP networks must convert these analog signals into digital packets before carrying them over the IP network. Once transported, the signals are recreated into sound waves for your listening pleasure.

Transporting Packets in Real Time
Because of the near worldwide existence of circuit- switched telephony, VoIP will always be compared to its circuit-switched predecessor. The sound quality of VoIP is based on the network’s ability to deliver packets with a high success rate (more than 99 percent) and minimal delay (less than 150 msec end to end). There are well-established standards for quality. Although some people using VoIP might be willing to tolerate lower-quality sound in exchange for free long distance, VoIP quality must match that of circuit-switch telephony in business applications.

Analog Voice to Digital Conversion
Traditional telephony systems convert the sound waves produced by the human voice to electrical signals, which are easily transmitted down a wire. On the receiving end of the connection, the electric signals excite a diaphragm, which produces a very good representation (or an analogy) of the original signal.

Analog signals consist of continuously variable waveforms having an infinite number of states; therefore, you can theoretically replicate them exactly. Digital telephony (including VoIP) must convert the original signal to a digital stream (or series of packets) on the transmitter and then recreate it on the receiving end.

Analog to digital conversion (A/D) happens through sampling.

Sampling is the process of taking instantaneous measurements of an analog signal. If you take enough samples, you can replicate the original analog signal by "connecting the dots" of the instantaneous measurements.

To correctly replicate the original signal, you must take the proper ratio of samples. If you take too few samples, more than one signal (frequency) can connect the dots. This process is called aliasing. Taking too many samples, however, is not always better. Over-sampling can improve the accuracy of the replicated signal, but at some point, it consumes too many resources (CPU and bandwidth) without yielding additional benefits.

It turns out that the ideal sampling rate for any signal is twice that of the signal’s highest frequency. In other words, you can accurately recreate a 2 cycles per second signal by sampling it 4 times per second. This rate is called the Nyquist rate, named after the AT&T engineer who discovered it. The Nyquist theorem states that you can digitally recreate any analog signal by sampling it at twice the rate of the highest frequency contained in the signal. Typically, devices sample signals at just over the Nyquist rate.

One of key issues with VoIP is the conservation of bandwidth. Because the routing information contained in VoIP packets can more than double the size of the packet, it is important to compress the voice data as much as possible. Compression has three levels, or orders. The first order is to simply not transmit what can’t be heard. A typical conversation is mostly silent (hard to believe but true). These silent parts of speech are not transmitted.

The second order of compression is to get the most out of the digital conversion of the analog signal. Remember that the analog signal has infinite number of states, but the digital representation must be a series of is and Os and is limited by the number of bits used. More bits means more levels (a good thing) and more bandwidth required (a bad thing). For example, an 8-bit digital signal could represent 256 levels. Any instantaneous measurement in the A/D process is represented by one of these levels. This process is referred to as quantization. It turns out that by stacking more levels at low amplitudes (rather than evenly spacing them Out), you can use fewer bits to get the same quality you would get by using more levels (without consuming additional bandwidth).

Finally, the third order is to not send the actual voice data. You can model speech signals using pitch and tone. There are wide variances of tone and pitch data, but you can store them in lookup tables. Modern computing techniques and impressive statistical modeling let you send the table location (or vectors) of the pitch and tone information across the network. The receiving end applies the vectors to the tables and recreates the sound.

Comfort Noise
The digital signals in VoIP are usually much "cleaner" than the analog signals used in circuit-switch telephony. In analog systems, any amplification of the signal also amplifies noise, resulting in static heard in the background during calls. Digital signals can clean out noise and achieve a purer sound. This improvement might seem like a good thing, but it actually causes problems.

It turns out that on analog calls, the slight background noise is an indication of a good connection. Because most of the world’s phone users were trained to hear this noise, the absence of the static bothers people and makes some wonder if the connection is still live ("Hello?"). To mitigate this "problem," digital systems inject static on the receiving end to let users know they still have a good connection. This injected static is called comfort noise.