Voice activity identification for speaker tracking in a packed based conferencing system with distributed processing

Abstract

A distributed conferencing system has a plurality of conferencing nodes to connect groups of participants to a conference. Each of the conferencing nodes provides for the connection of one or more participants to the conference. Each node includes a DSP for distributed signal processing. The node DSP includes: A signal measuring device for measuring features of the signals from each of the participants, such as power, zero crossing rate and short term energy. The nodes include voice activity determination and a communication device for communicating the measured signal characteristics for a plurality of participant input signals to all other conferencing nodes. Muting means for muting individual participant input signals so that only selected signals are transmitted over the conference bus to the other participants. The voice activity detection utilizes a state machine with three states, voice state, transition state and noise state, dependant upon the measured energy level, zero crossing rate and other features of the signals. A high threshold and a low energy threshold; zero crossing rates; average energies; energy level means and variances and other features are used in differentiating voice and noise. The state machine will not move directly from voice to noise state but will move to a transition state first, to reduce the likelihood of missclassification of a weak voice signal as noise and to avoid frequent clipping which can be caused if the state machine moves to noise state during brief pauses in voice.

Dependencies are inferred from the claim text.