A method and apparatus for performing a active packet bundling in a VoIP (Voice over Internet Protocol) based on a measure of voice concealability of one or more speech packets to be transmitted. Voice concealability is a property of a speech packet which indicates the likelihood that a given speech segment will meet the “quasi-stationary” (QS) behavior assumed by conventional Packet Loss Concealment (PLC) algorithms used to compensate for transmission loss of speech packets. By ensuring that only well-concealed frames are delayed for bundling, it is more likely that the receiver will run the PLC algorithm when it is most effective (i.e., when it is likely to perform well at reconstructing the missing speech frame), rather than when it is more likely to be ineffective. This allows for a reduction in bandwidth while minimizing any potential negative impact on quality.