Machine Learning-based Time-slot Time-varying Filtering for Mandarin Tone Processing

Abstract

Tone is one of the biggest differences between Mandarin as a tonal language and English, meaning that changes in frequency will appear in every character and determine their meaning. Therefore, the accurate identification of intonation plays an important role in Chinese language processing and learning. In the learning and training process of Mandarin Chinese, graphical feedback can provide valuable assistance to linguists or learners, providing them with more insight than simple accuracy ratings as they can determined what to change to improve their intonation based on this feedback. In this paper, we refer to the process of generating Mel-frequency cepstrum coefficients and employ a machine learning-aided time-slot filtering approach to obtain an intuitive transformation of pitch from Mel spectrum to geometric graphs as feedback to Chinese language workers and learners, and to help them improve their mastery of Mandarin tones. Our preliminary results show that these illustrations can accurately represent frequency changes in relation to time. At the same time, the lower computational cost of the optimization allows it to be used as learning software on various mobile platforms.