Adapting Multilingual Neural Network Hierarchy to a New Language

feature extraction, Bottle-Neck features, neural network adaptation, multilingual neural networks, Stacked Bottle- Neck structure

This work addresses a thorough evaluation of multilingual techniques for adapting feature extraction neural network hierarchy - Stack Bottle-Neck system.

The neural network based features became an inseparable part of state-of-the-art LVCSR systems. With the increasing accent on fast development of ASR system on limited resources, there is an effort to alleviate the need of large amount of transcribed in-domain data. One successful way is to use data from other languages. We present extensive evaluation of several strategies to adapt hierarchical neural network in search for the most effective one. To avoid the bias towards one target language, our strategies were evaluated on five languages. Also, several multilingual neural network hierarchies were trained on two sets of languages. Thus the results provide solid insight into the problem of adapting hierarchical system.