A new score normalization for text-independent speaker verification

Abstract

In iVector-based speaker verification system, the claimed speaker was verified if the similarity between the iVector of the tested utterance (iVector-ts) and the iVector of the claimed speaker (iVector-cs) is smaller than a fixed threshold. The commonly used method to measure the similarity between the iVector-ts and iVector-cs is the cosine similarity scoring method. To further improve the performance of the speaker verification system when the training data is insufficient, a new scoring method termed as ratio normalization (Rnorm) scoring method is proposed, where the similarity between iVector-ts and iVector-cs is normalized by the dissimilarity between the tested speaker model and the universal background model (UBM). Preliminary experimental results with Timit database and self-built database show that our proposed Rnorm scoring method is able to reduce the equal error rate (EER) of the iVector-based TIV speaker verification system compared with that of using conventional cosine similarity scoring method.