A Data-driven Approach to Speech/Non-speech Detection

We present a data-driven approach to weighting the temporal context of signal energy to be used in a simple speech/non-speech detector (SND). The optimal weights are obtained using linear discriminant analysis (LDA). Regularization is performed to handle numerical issues inherent to the usage of correlated features. The discriminant so obtained is interpreted as a filter in the modulation spectral domain. Experimental evaluations on the test data set, in terms of average frame-level error rate over different SNR levels, show that the proposed method yields an absolute performance gain of $10.9\%$, $17.5\%$, $7.9\%$ and $8.3\%$ over ITU's G.729B, ETSI's AMR1, AMR2 and a state-of-the-art multi-layer perceptron based system, respectively. This shows that even a simple feature such as full-band energy, when employed with a large-enough context, shows promise for applications.