If the symptom stops, then comfort noise generation is the likely cause of the problem. Reducing the music-threshold at which voice is detected or increasing voice vad-time values on the gateway might make the hissing or clipping less noticeable without the need to disable VAD permanently.
These techniques essentially disable VAD at low volume levels or during small gaps, respectively. It is not practical to just disable comfort noise, as that causes other voice quality symptoms like clicking or gaps of absolute silence between sentences.
If these tuning techniques do not solve the problem, then VAD needs to be disabled, with a resulting loss in bandwidth savings.
These are the reasons why VAD can cause these issues:
The use of VAD to free transmission bandwidth requires the end-point either to play out silent (zero) samples or to play out a representation of the noise from the other end.
Playing silent samples is a simple solution. However, it has serious voice quality impacts. These include choppiness with short pauses and a dead-air effect during long pauses.
Representations of far-end noise can be done in a variety of ways. For example, if a call is made from a Cisco 7960 to another Cisco 7960, the transmitting device sends Silence Information Descriptor (SID) frames that enable the comfort noise generator on the receiving device to generate the proper level of comfort noise. Information present in the original background noise that was dropped by the silence detector (possibly low-level speech or music) is lost at the far-end. If the transmitting device is not a Cisco 7960, the receiving 7960 approximates the appropriate comfort noise from previously received packets.
The ITU has defined a VAD standard, called G.729B, which only works with G.729 and G.729a (8k) compression. The ITU implementation is optimized for G.729 and G.729a and it also reduces the cumulative delay of G.729 and G.729a and VAD by doing the voice compression and VAD processing in parallel. According to a paper published in the IEEE Communications magazine, the Mean Opinion Score (MOS) scores of G.729B are comparable to the non-VAD G.729 and G.729a case. The G.729B also has the benefit of being an ITU standard.
As with many bandwidth conservation mechanisms, VAD and comfort noise have trade-offs. In particular, artifacts such as clipping of soft voice segments, loud (or artificial) comfort noise (hissing) and additional delay might be experienced. If the device combination or the background noise characteristics for certain calls cause comfort noise to be annoying, then the only fool-proof way to currently avoid this problem is to disable it. Usually, disabling comfort noise also necessitates disabling VAD.