Robustness of speaker normalization approaches: A study
Rohit Sinha, B. Sandeep Kumar · 2008
In this work, we have studied the effectiveness of recently proposed warp factor estimation method based on correlation between the pitch frequency and the warp factor for speaker normalization. Our study shows that estimation of warp factors using maximum a posteriori criterion by exploiting the statistics from the pitch based warp factor estimation approach results in 12.5% relative improvement in performance compared to conventional maximum likelihood based method. The exploitation of statistics from pitch based approach is also found to reduce the search complexity. We have also studied the noise robustness of the above mentioned speaker normalization methods under different types of additive noise conditions.