Improvement of distant-talking speaker identification using bottleneck features of DNN

In this paper we propose bottleneck features of deep neural network for distant-talking speaker identification. The accuracy of distant-talking speaker recognition is significantly degraded under reverberant environment. Feature mapping or feature transformation has been shown efficacy in channel-mismatch speaker recognition. Bottleneck feature derived from multilayer network, which is a nonlinear feature transformation method, has been shown efficacy in automatic speech recognition (ASR) system. In this study, bottleneck features extracted from deep neural networks (DNNs) which employ an unsupervised pre-training method are used as nonlinear feature transformation for distant-talking speech. The speaker identification experiment was performed on large-scale distant-talking speech set, with reverberant environments different to the training environments. The proposed bottleneck features achieved a relative error reduction of 46.3% compared with conventional MFCC. Moreover, a combination of likelihoods of bottleneck features and MFCC achieved a furthermore improvement.