An Approach for Formant Based Speech Recognition in Noise

In this paper, a noise robust formant frequency estimation scheme is developed utilizing the advantageous properties of the autocorrelation function of the band-limited noisy speech signal. It is shown that the use of autocorrelation operation on a speech signal, which is band-limited to a particular formant zone, in comparison to one without any band limitation, can provide higher noise immunity, especially under severe noisy condition. In order to extract each formant, a modified higher order Yule-Walker method is employed on the resulting autocorrelation sequence. Within a band, the pole with the maximum energy is selected as the formant. The estimated formants are used as features along with conventional Mel frequency cepstral coefficients in a vowel recognition system, where the linear discriminant based classifier is utilized. Extensive experimentation is carried out on speech samples taken from the TIMIT standard speech database. It is found that the proposed algorithm provides superior formant estimation accuracy in comparison to that obtained by some of the state of the art methods even at a very low level of signal-to-noise ratio (SNR) for both male and female speakers. Moreover, formant estimates obtained by the proposed method can also provide better vowel recognition accuracy in the presence of significant background noise.