Environmental robust features for speech detection
Thomas Kemp, Climent Nadeu, Y.H. Lam, Josep Maria Sola i Caros · 2004
In this paper, two novel features, Line Spectrum Center Range and Line Spectrum Flux, both derived from Line Spectrum Frequencies, are proposed to detect the presence of speech in various acoustic environments. Evaluation results using Fischer Discriminant Analysis and Scatter Matrices indicated that the new features excel the state-of-theart features. An environmental robust hybrid feature set including the proposed features, Normalized Energy Dynamic Range and Mel-Frequency Cepstrum Coefficients is further introduced. When evaluating the hybrid feature set on a Gaussian Mixture Model based classification engine, the results showed that the hybrid feature set outperformed MelFrequency Cepstrum Coefficients up to in terms of relative frame error rate.