Semi-development and Performance Evaluation of a Pronunciation Judgment System Using Free Machine Learning Services

Abstract

With the recent development of the global community, communication skills in languages other than one's native tongue have become essential. While attending school is an effective method of language learning, it is often difficult due to time and financial constraints. If it becomes possible to acquire language pronunciation through self-study, it is expected that the learning process can be significantly shortened. Furthermore, building such a practice system without specialized IT knowledge, such as programming, would be a great benefit to educational settings. This paper describes a pronunciation assessment system for language learning that utilizes a free machine learning service. The target language is Japanese. By providing machine learning with speech data of homonyms that are difficult for non-native speakers to distinguish, we build a system that can assess the accuracy of pronounced words. For machine learning, we use Google Teachable Machine, a free service that allows system building without specialized IT knowledge. Experiments using this method demonstrate that we have constructed a system that can assess the accuracy of native speaker pronunciation with a very high probability.