Code-switching Speech Detection Method by Combination of Language and Acoustic Information

Abstract

In this paper, we propose a new speech detection method to English-Mandarin code-switching speech. Unlike previous methods, in this method we first train a support vector machine (SVM) model based on feature parameters and Gaussian Mixture Model (GMM) , then integrate the language identification (LID) information based on SVM model and acoustic information into the decoding process. Lastly, we develop a prototype system to present the method. Experiments proved that our method we can improve the accurancy of code-switching speech recognition at a certain degree compared with previous methods.

Language vs Speaker Change: A Comparative Study

“…Humans may exploit the long term spectro-temporal relations to discriminate between languages. Motivating from the same, SLD system can also be developed using acoustic-phonetic information exploited at suprasegmental levels (Liu et al, 2021;Zhang, 2013). The advantage of the acoustic-phonetic based approach is, it does not require transcribed speech data.…”