Question: Is echo inherent to VoIP? Is it possible to eliminate it? What is the minimum bandwidth requirement to guaranty a good voice quality all the time?
Answer: There are two main sources of echo in telephony networks: line echo and acoustic echo.
Acoustic echo is generated on any phone (IP or otherwise) when there is feedback from the speaker to the microphone. This is particularly noticeable on many speaker phones.
Line echo is very common in the PSTN network and this most commonly occurs when there is a two wire to four wire conversion in the network (for example, where analog is converted into T1 or E1).
To combat these types of echo, there are echo cancellers. As you can probably imagine, there are acoustic echo cancellers (AEC) and line echo cancellers (LEC). How well the echo is cancelled depends upon the quality of the echo canceller.
One key parameter in an echo canceller is the tail length. Basically the way an echo canceller works is it remembers the waveform sent out, and for a certain period of time looks to see a waveform coming back that it can correlate to the original signal (usually arriving later, at lower amplitude, and with more noise). Typically, echo cancellers can be set to 32ms, 64ms, or 128ms tail lengths. If the return signal (echo) arrives too late, the echo canceller won't be able to properly correlate and cancel it. In summary, it is possible to greatly minimize or nearly eliminate echo if proper echo cancellation is in place.
If you're using an uncompressed G.711 codec, over Ethernet, you need approximately 87 kb/s in each direction to carry on a conversation. If you are using a compressed codec such as G.729, you need approximately 24 kb/s in each direction. This is the minimum bandwidth required to carry on a conversation. Note that bandwidth alone does not guarantee good voice quality. If there are dropped packets, random delays or other things of this nature, the voice quality may not be good. You need to have a properly designed network to ensure decent voice quality. Congestion points should be eliminated. If there is going to be traffic congestion, a quality of service mechanism that prioritises voice traffic over other traffic should be used.