Tracking of vocal tract resonances based on dynamic programming and Kalman filtering

In this work we propose a method for accurate and reliable estimation of vocal tract resonance (VTR) frequencies. The proposed method is based on finding candidates of the VTR frequencies using LPC analysis and estimation of the VTR using dynamic programming and Kalman filtering/smoothing. The performance of the proposed method is measured in terms of recently proposed hand-labeled MSR-VTR database (Li Deng et al., 2006). The performance of the proposed method is compared with two different formant tracking computation methods. One way is that we compare the results from proposed method with the results from output dynamic programming algorithm, which is our baseline system. The second way is that we compare the proposed method results with the results of WaveSurfer. The proposed method reduces the vocal tract resonance estimation error rate 13%, over the baseline system and 29% over the WaveSurfer.