Are you sure that there aren’t a lot of handovers (for example due to Quality) and you are in one cell?
As you know durring handover speech is substituted using some algorithm with last good received speech frame. In some vendors it is the same one but each next is more silenced and after certain time it is only silence.
Otherwise is sounds as equipment problem. Noise reduction, echo cancelers (GSM-GSM), transcoders etc.