Parametric Mixing for Centralized VOIP Conferencing using ITU-T Recommendation G.722.2

VoIP conferencing with a centralized speech mixing bridge introduces additional end-to-end latency into packetized voice communication. This paper investigates how full tandem speech decoding, time-domain mixing, speech encoding cycle can be circumvented by instead extracting the coded speech parameters and performing the speech packet mixing without time-domain reconstruction. By mixing through coded speech parameters, we show that nearly an 85 % decrease in computational complexity can be achieved over full tandem mixing of two speakers for G.722.2, thus significantly reducing the packet latency at the centralized speech mixing bridge. For the G.722.2 parametric mixer presented, linear prediction coefficients (LPCs), pitch lags, fixed codebooks, and gains, are extracted (without full speech reconstruction) from the encoded bit stream, mixed, and then re-encoded instead of the full tandem approach where each speech frame must be fully reconstructed. We investigate the mixing in two scenarios: i) mix two 12.65 kbps G.722.2 speech streams at a mixed rate of 12.65 kbps, and ii) mix two 12.65 kbps G.722.2 speech streams at a mixed rate of 18.25 kbps. PAMS is used to evaluate the speech quality of the parametric mixer, resulting in an average distortion 0.37 MOS (compared to tandem mixing) as shown by simulations using typical conversation models