Acoustic model training using feature vectors generated by manipulating speech parameters of real speakers

In this paper, we propose a robust speaker-independent acoustic model training method using generative training to generate many pseudo-speakers from a small number of real speakers. We focus on the difference between each speaker's vocal tract length, and manipulate it in order to create many different pseudo-speakers with a range of vocal tract lengths. This method employs frequency warping based on the inverted use Vocal Tract Length Normalization(VTLN). Another method for creating pseudo-speakers is to vary the speaking rate of the speakers. This can be achieved by a method called PICOLA; Pointer Interval Controlled OverLap and Add. In experiments, we train acoustic models using these generated pseudo-speakers in addition to the original speakers. Evaluation results show that generating pseudo-speakers by manipulating speaking rates did not result in a sufficient increase in performance, however, vocal tract length warping was effective.