使用深度信念網路的標準國語語調分析;Intonation Analysis in Standard Mandarin using Deep Belief Networks

[[abstract]]深度信念網路在機器學習領域是一個常用的方法，像是在圖像辨識、語音辨識等等。我們只需要對原始資料作些許的前處理，他就能夠從原始資料中學習特徵。在全世界的許多語言中，普通話是少數幾種聲調語言之一，不同聲調會讓一個語音有著不同的意思，也因此在語音辨識上普通話較其他語言辨識的困難度較高，例如英文。換句話說，聲調辨識的正確率對語音辨識率有著相當大的影響，在本論文中，使用了深度信念網路作為國語語調分析的工具，對連續語音進行聲調的分析，並與以往實驗的成果進行比較，實驗結果顯示使用深度信念網路作為工具能得到比以往的方法還要好的成果。 Deep belief network is a commonly used method in machine learning domain, such as image recognition, speech recognition, etc. With a little pre-processing, it can learn the features from the raw data excellently.In many languages around the world, Standard Mandarin is one of few tonal languages, different tones will make a speech differently. Therefore speech recognition in Standard Mandarin is highly difficult than other language for instance English. In other words, it means that the correctness of tonal recognition has a great influence on the correctness of speech recognition,In this thesis, Deep belief network is used as the tool for Intonation Analysis of Standard Mandarin, and the analysis of continuous speech is carried out. The results show that using Deep belief network as a tool can obtain more information than previous methods