Exploring the Role of Phonetic Bottleneck Features for Speaker and Language Recognition

MITCH MCLAREN; LUCIANA FERRER; AARON LAWSON

Shangai

Congreso; IEEE Conference on Acoustics, Speech and Signal Processing 2016; 2016

IEEE

Using bottleneck features extracted from a deep neural network(DNN) trained to predict senone posteriors has resulted in new,state-of-the-art technology for language and speaker identification.For language identification, the features? dense phonetic informationis believed to enable improved performance by better representinglanguage-dependent phone distributions. For speaker recognition,the role of these features is less clear, given that a bottleneck layernear the DNN output layer is thought to contain limited speaker in-formation. In this article, we analyze the role of bottleneck featuresin these identification tasks by varying the DNN layer from whichthey are extracted, under the hypothesis that speaker information istraded for dense phonetic information as the layer moves toward theDNN output layer. Experiments support this hypothesis under cer-tain conditions, and highlight the benefit of using a bottleneck layerclose to the DNN output layer when DNN training data is matchedto the evaluation conditions, and a layer more central to the DNNotherwise.