MagNetO: X-vector Magnitude Estimation Network plus Offset for Improved Speaker Recognition
Daniel Garcia-Romero, Greg Sell, Alan Mccree |
---|
We present a magnitude estimation network that is combined with a modified ResNet x-vector system to generate embeddings whose inner product is able to produce calibrated scores with increased discrimination. A three-step training procedure is used. First, the network is trained using short segments and a multi-class cross-entropy loss with angular margin softmax. During the second step, only a reduced subset of the DNN parameters are refined using full-length recordings. Finally, the magnitude estimation network is trained using a binary cross-entropy loss over pairs of target and non-target trials. The resulting system is evaluated on 4 widely-used benchmarks and provides significant discrimination and calibration gains at multiple operating points.