Voice Activity Detection Using Wavelet-Based Multiresolution Spectrum and Support Vector Machines and Audio Mixing Algorithm

This paper presents a Voice Activity Detection (VAD) algorithm and efficient speech mixing algorithm for a multimedia conference. The proposed VAD uses MFCC of multiresolution spectrum based on wavelets and two classical audio parameters as audio feature, and prejudges silence by detection of multi-gate zero cross ratio, and classify noise and voice by Support Vector Machines (SVM). New speech mixing algorithm used in Multipoint Control Unit (MCU) of conferences imposes short-time power of each audio stream as mixing weight vector, and is designed for parallel processing in program. Various experiments show, proposed VAD algorithm achieves overall better performance in all SNRs than VAD of G.729b and other VAD, output audio of new speech mixing algorithm has excellent hearing perceptibility, and its computational time delay are small enough to satisfy the needs of real-time transmission, and MCU computation is lower than that based on G.729b VAD.