A study of phonetic feature representations for SVM-based speaker verification

摘要

We investigate an alternative formulation of phonetic feature representations for SVM-based speaker verification. The new features are based on conditional likelihood representations rather than the joint-likelihood or bag-of-ngram calculations traditionally used. Conditional likelihoods are shown to be a more natural method of modelling phonetic information, and improve upon conventional joint likelihoods in a number of cases. The problem of feature normalisation is also examined, with a previously proposed non-parametric method based on rank shown to be particularly useful. Combinations of feature representations are examined and the potential for complementary information between joint and conditional likelihoods considered. Additionally, feature compensation is applied to conditional likelihoods with considerable improvement in performance.