Across-phone variability and diagonal term in joint factor analysis for speaker recognition

Abstract

We investigate usefulness of across-phone variability for speaker recognition in a joint factor analysis (JFA) framework. We estimate the variability as across-phone covariance within a conversation side averaged over all conversations. Note that it is a part of channel variability in the current JFA framework. We independently estimate feature subspaces representing across-phone, speaker and channel variability and perform speaker recognition experiments by either keeping them or removing them. The results show that the across-phone subspace is more correlated with the speaker subspace. We also perform speaker recognition experiments when combining the subspaces. Results show an improvement when phone and speaker subspaces are combined. This shows that across-phone variability is useful for speaker recognition. Further experiments show that the results are affected by a diagonal term from JFA. In particular, the improvement when combining the speaker and phone subspaces is reduced when the diagonal term is estimated from a universal background model (UBM). This implies that there is an interaction between the variability represented by the diagonal term and the across-phone variability. Overall, the work shows the importance of understanding the diagonal term (with speaker and channel subspaces) for incorporating additional variability into JFA beyond speaker and channel.

Keywords— Speaker recognition, joint factor analysis, phonetic variability, language independent speech recognition