Formants are useful for phoneme detection which is paramount for speech recognition. Though it is relatively easy to find formants for single utterances, formant tracking is more challenging. Because speech is not uncorrelated across time, Linear Predicitice Codes have been used for formant tracking. Here, study the application of Extended Kalman Filter for formant tracking in comparison with other methods.

We studied methods based on Linear Predictive Codes (LPC) for formant tracking in speech signal. In these methods, the vocal filter is assumed to be an all pole filter and the coefficients are estimated from the signal value. The performance of this method degrades if the number of polesis not chosen properly. Extended kalman filter performs better than LPC if the matrices that capture model parameters are estimated correctly. These methods' performances were compared with the formant tracker in PRAAT and the advantages and pitfalls of the methods were highlighted. In particular, the EKF method can track vowels, plosives and fricatives better than LPC. Also, EKF method continues tracking after silence much better than LFC.