Distributed automatic text-independent speaker identification using GMM-UBM speaker models

Abstract

The ETSI “Aurora” is a digit-based standard developed for distributed speech recognition (DSR) over telephone communication channels. This paper introduces a digit-based text-independent distributed speaker identification (DSID) system over telephone channels within the DSR framework. In this DSID system, the hypothesized speaker model is derived by GMM-UBM model training using Aurora2 connected digit training speech data and maximum a posteriori (MAP) adaptation. The UBM technique for speaker models is incorporated into this DSID system to reduce the computational complexities significantly. Experiments on the Aurora2 speech recognition corpus show that GMM-UBM yields excellent performance for speaker recognition over telephone channels. Compared to the baseline system, we got 100% recognition accuracy for this proposed DSID within the ETSI DSR framework.