INHIBITING REFERENCE PATTERN GENERATING METHOD

摘要

The present invention relates to inhibiting reference pattern generating methods used in speaker verification systems. Input speech is supplied to an analyzer for conversion to feature vector, which is supplied to an inhibiting speaker selector and also to a feature vector memory. An inhibiting speaker selector selects inhibiting speakers by calculating similarities between reference patterns stored in an inhibiting speaker's reference pattern memory and the feature vector of the true speaker, stored in the feature vector memory. The inhibiting speakers are selected in the order of smaller distances thereof. Selected inhibiting speaker's reference patterns are supplied to an estimator. The estimator obtains mapping functions from inhibiting speaker's feature vector spaces to feature vector space of the true speaker, using the reference patterns supplied from the inhibiting speaker selector and the feature vector of the true speaker, supplied from the feature vector memory. The mapping functions thus obtained are supplied to an adapter. The adapter adapts the inhibiting speaker's reference patterns selected in the inhibiting speaker selector by using the mapping functions supplied from the estimator.