Bottleneck and Embedding Representation of Speech for DNN-based Language and Speaker Recognition

Abstract

In this manuscript, we summarize the findings presented in Alicia Lozano Diez's Ph.D. Thesis, defended on the 22nd of June, 2018 in Universidad Autonoma de Madrid (Spain). In particular, this Ph.D. Thesis explores different approaches to the tasks of language and speaker recognition, focusing on systems where deep neural networks (DNNs) become part of traditional pipelines, replacing some stages or the whole system itself. First, we present a DNN as classifier for the task of language recognition. Second, we analyze the use of DNNs for feature extraction at frame-level, the so-called bottleneck features, for both language and speaker recognition. Finally, utterancelevel representation of the speech segments learned by the DNN (known as embedding) is described and presented for the task of language recognition. All these approaches provide alternatives to classical language and speaker recognition systems based on i-vectors (Total Variability modeling) over acoustic features (MFCCs, for instance). Moreover, they usually yield better results in terms of performance.

Unsupervised Regularization of the Embedding Extractor for Robust Language Identification

“…Our main contribution is to perform this adaptation for the segment-level embedding extractor. The segment-level embedding extractor of a state-of-the-art language identification system is a neural network [11], which can be used for direct classification [12], or just to extract representations [4] which are then processed by a classifier. Different model-based domain adaptation methods have been introduced in the neural network training literature.…”