一個適用於聲調語言的辨認器

Abstract

在本文中，我們將介紹一個適用於聲調語言的辦認器。在語音辦認中，通常只有短時間的頻譜特性被使用，諸如線性預估編碼係數（LPC Coefficients，），倒頻譜係數（Cepstral Coefficients），頻帶係數（Filter Banks）等等，但這些參數對於聲調語音辦認是不夠的，因為還存在一個非常重要的聲韻特性，那就是聲調。在本文中，我們組合了對於短時間頻譜特性以及聲調特性的距離量測，而產生一個適用於聲調語言的辦認器，在我們的實驗中，所用的語彙是國語0至9的數字共有100個人，50男和50女參予我們的實驗，在不考慮聲調特性的情況下，對於測試者不需預先訓練語音的實驗，辦認率只有91.0%，如果聲調的特性加以考慮，則辦認率提高到95.9%。 In this paper, we present a recognizer for tone languages. In general, only the short-time spectral features, like LPC coefficients, cepstral coefficients, and filter banks, etc. are used in speech recognition. But these features are insufficient for speech recognition in tone languages, because another distinctive feature, tone, also plays a significant role in the phonological system of tone languages. Here we combine the distance for short-time spectra and the distance for tone to form a recognizer for tone languages. The experimental data are 0-9 digits in Mandarin. A total of 100 informants, 50 male and 50 female, are involved. If the effects of tone are neglected, the average accuracy in a speaker-untrained test is 91.0%. If the tone’s distance is included, the accuracy can be promoted to 95.9%.