Standoff speaker recognition: effects of recording distance mismatch on speaker recognition system performance

Abstract

Speech can potentially be used to identify individuals from a distance and contribute to the growing effort to develop methods for standoff, multimodal biometric identification. However, mismatched recording distances for the enrollment and verification speech samples can potentially introduce new challenges for speaker recognition systems. This paper describes a data collection, referred to as the Standoff Multi-Microphone Speech Corpus, which allows investigation of the impact of recording distance mismatch on the performance of speaker recognition systems. Additionally, a supervised method for linear subspace decomposition was evaluated in an effort to mitigate the effects of recording distance mismatch. The results of this study indicate that mismatched recording distances have a consistent negative impact on performance of a standoff speaker recognition system; however, subspace decomposition techniques may be able to reduce the penalty observed with mismatched recording distances.