Investigation of Frame Alignments for GMM-based Text-prompted Speaker Verification
The frame alignment acts as an important role in GMM-based speaker verification. In text-prompted speaker verification, it is common practice to use the transcriptions to align speech frames to phonetic units. In this paper, we compare the performance of alignments from hidden Markov model (HMM) and deep neural network (DNN), using the same training data and phonetic units. We incorporate a phonetic Gaussian mixture model (PGMM) in the DNN-based alignment, making the total number of Gaussian mixtures equal to that of the HMM. Based on the Kullback-Leibler divergence (KLD) between the HMM and DNN alignments, we also present a fast and efficient way to verify the text content. Our experiments on RSR2015 Part-3 show that, even with a small training set, the DNN-based alignment outperforms HMM alignment. Results also show that it is not effective to utilize the DNN posteriors in the HMM system. This phenomenon illustrates that under clean conditions (e.g., RSR2015), text-prompted speaker verification does not benefit from the use of transcriptions. However, the prompted text can be used as pass-phrase to enhance the security. The content verification experiment demonstrates the effectiveness of our proposed KLD-based method.
READ FULL TEXT