Speech event detection using SVM and NMD

In this paper we propose a speech event detector that segments speech signals in terms of four broad acoustic-phonetic classes of events. Frame-based detection was carried out using support vector machines (SVM). Non-negative matrix deconvolution (NMD) was used in order to switch from a frame-based detection to a segment-based detection. Results obtained using the TIMIT corpus are reported and compared to a broad class detector based on hidden Markov models (HMM) with a MFCC front-end. It was found that the proposed SVM/NMD system outperforms the HMM system in what concerns to accuracy and also to the quality of he detected boundaries.