In this paper, we propose a data augmentation scheme with bandwidth extension (BWE) for deep neural network (DNN)-based automatic speaker verification (ASV) systems. One of the DNN-based ASV systems which is named "x-vector" requires a large amount of training data. Especially, using a large amount of wideband (WB) data obtains one of the highest performances for the x-vector-based ASV systems. However, when amount and variety of data are limited, it is important to use data augmentation schemes. If the BWE methods can use as data augmentation schemes for x-vector-based systems, the issue on amount and variety of data is relaxed. Some reports have already considered using extended WB data from narrowband (NB) data by DNN-based BWE. Recently, the authors have reported that the effectiveness of BWE methods for machine learning frameworks. Additionally, the quality of generated speeches by non-learning-based BWE is almost same as learning-based BWE. Therefore, in this paper, we aim to demonstrate several non-leaning-based BWE methods are useful as data augmentation for x-vector-based ASV systems. By using Speakers In The Wild database and NIST SRE one, experimental results showed that the proposed system provided the error reduction of 22.7%, compared with our baseline system.
Cite as: Miyamoto, H., Shiota, S., Kiya, H. (2020) Application of Bandwidth Extension with No Learning to Data Augmentation for Speaker Verification. Proc. The Speaker and Language Recognition Workshop (Odyssey 2020), 446-450, doi: 10.21437/Odyssey.2020-64
@inproceedings{miyamoto20_odyssey,
author={Haruna Miyamoto and Sayaka Shiota and Hitoshi Kiya},
title={{Application of Bandwidth Extension with No Learning to Data Augmentation for Speaker Verification}},
year=2020,
booktitle={Proc. The Speaker and Language Recognition Workshop (Odyssey 2020)},
pages={446--450},
doi={10.21437/Odyssey.2020-64}
}