Robust tracking of glottal LF-model parameters by multi-estimate fusion

A new approach to robust tracking of glottal LF-model parameters is presented. The approach does not rely on a new glottal source estimation algorithm, but instead introduces a new extensible multi-estimate fusion framework. Within this framework several existing algorithms are applied in parallel to extract glottal LF-model parameter estimates which are subsequently passed to quantitative data fusion procedures. The preliminary implementation of the fusion algorithm described here incorporates three glottal inverse filtering methods and one time-domain LF-model fitting algorithm. Experimental results for both synthetic and natural speech signals demonstrate the effectiveness of the fusion algorithm. The proposed method is flexible and can be easily extended for other speech processing applications such as speech synthesis, speaker identification and prosody analysis.