A Feature for Voice Activity Detection Derived from Speech Analysis with the Exponential Autoregressive Model
Kentaro Ishizuka, Hiroko Kato · 2006
This paper proposes a feature for voice activity detection (VAD) obtained from a speech signal analysis that uses the exponential autoregressive (ExpAR) model. This model employs exponential terms that depend on the amplitude of observed signals in the AR coefficients part. Since these terms can model the nonlinearity of speech caused by the nonlinear fluctuation of vocal cord vibration, this model can provide a better fit for speech signals. A parameter in the exponential terms of the ExpAR model called 'the scaling parameter,' is directly associated with the degree of nonlinearity of analyzed signals. Therefore, the scaling parameter changes when observed signals include speech signals. Based on this property, this parameter is usable as a feature for VAD under noisy conditions. An experiment using noisy speech data confirmed the potential performance of the proposed feature by comparing receiver operating characteristics curves obtained from the proposed feature and conventional robust features. Another experiment was conducted by comparing recalls, precisions, and F-measures for speech interval detection achieved by our proposed VAD algorithm, that utilized only the proposed feature, and two widely used standardized algorithms. The result showed that the proposed method could achieve better performance than those of the standardized algorithms