The purpose of this study is to involve both Convolutional Neural Networks and a typical learning algorithm in the allophone classification process. A list of words including aspirated and non-aspirated allophones pronounced by native and non-native English speakers is recorded and then edited and analyzed. Allophones extracted from English speakers’ recordings are presented in the form of two-dimensional spectrogram images and used as input to train the Convolutional Neural Networks. Various settings of the spectral representation are analyzed to determine adequate option for the allophone classification. Then, testing is performed on the basis of non-native speakers’ utterances. The same approach is repeated employing learning algorithm but based on feature vectors. The archived classification results are promising as high accuracy is observed.

Authors: Piotrowska, Magdalena; Korvel, Grazina; Kurowski, Adam; Kostek, Bozena; Czyzewski, Andrzej
Affiliations: Vilnius University, Vilnius, Lithuania; Gdansk University of Technology, Gdansk, Poland(See document for exact affiliation information.)
AES Convention: 145 (October 2018) Paper Number: 10070
Publication Date: October 7, 2018 Import into BibTeX
Subject: Acoustics and Signal Processing
Permalink: https://www.aes.org/e-lib/browse.cfm?elib=19796

Download Now (626 KB)

This paper is Open Access which means you can download it for free.

Learn more about the AES E-Library

E-Library Location:   /conv/145/10070.pdf

Start a discussion about this paper!