Vocal tract length normalization using linear transformation based on maximum likelihood estimation

摘要

Vocal tract length normalization (VTLN) is one of the popular speaker adaptation techniques for speech recognition. The present study proposes a new VTLN algorithm in which expectation-maximization (EM) based parameter adaptation of HMM to vocal tract length is achieved in the mel-cepstral domain by utilizing a linear transformation model. Compared to other existing approaches based on hi-linear transformation for VTLN where a specific non-linear frequency warping function is employed in the spectrum domain and parameter adaptation of HMM is carried out in the cepstral domain, the proposed approach assumes a linear frequency warping with a single scaling factor and equivalent operation is modeled in the mel-cepstral domain by using a first order Taylor series approximation. The proposed scheme demonstrates significant improvement of recognition performance in a speaker independent word recognition task.