Features for phoneme independent speaker identification

This paper describes a unique cross-phoneme speaker identification experiment, using deliberately mismatched phoneme sets for training and testing. The underlying goal is to identify features that represent broad individually unique characteristics rather than those that represent phonetic differences, as are more typical of modern speaker identification and verification systems. A wide range of features are proposed and evaluated within this context using a Gaussian Mixture Model framework. The results show that log-area ratio has better phonetic independence than MFCCs, that residual phase carries substantial speaker information, and identifies several other features that also have usefulness for speaker identification independent of phonetic content.