使用統計模式之基頻軌跡偵測器

Abstract

本論文一開始用直覺的想法，運用語音基本的特性，對於每個音框找到可能的基頻值，再將高度接近的的基頻值連接成一個基頻區段並定義基頻區段的分數，區段與區段間，同樣也是利用合理的假設，將數個基頻區段連結成完整的 pitch contour。接下來從觀察中發現，這樣子的基本系統的確可以滿足大部分的情況後，在利用基本系統的結果，觀察什麼樣的特徵向量可以拿來判別voiced及unvoiced的音框，進而得到voiced/unvoiced的觀察機率模型；接下來還是運用機率模擬的方法，找到頻率變化由unvoiced到voiced、voiced到voiced、voiced到unvoiced的轉換機率，結合上述觀察機率模型以及轉換機率模型，我們就可以得到一個統計模式基頻偵測器，我們還觀察基頻區段間的變化，用機率模擬的方式取代原本unvoiced到voiced的轉換機率，最後設計了幾個實驗來描述這三個系統的表現 。

In this thesis, a heuristic pitch-tracking algorithm was first proposed. Four most probable pitch contour candidates were constructed under the constraint of pitch frequency change between frames and segments for the input utterance. And, a heuristic measurement was defined to decide the best pitch contour among the candidates. In this method, a smooth pitch contour can be found, which is not determined frame-bye-frame or using a smooth method for error correcting. Then, a new statistical pitch detection algorithm was proposed. In the proposed method, the probabilistic measures for U/V classification of a frame, the pitch transition probabilities between frames and pitch segments were properly modeled. And, the ML search could be used to find the most probable pitch contour of the speech utterance. After comparing the results of the two methods, the statistical pitch detection method will get better performance.