以分段量測的動態時間歸正法則改善混淆群集的辨認

Abstract

動態時間歸正法則在處理混淆群集，如英語的字母和數字，或是國語的單音節時，辦認品質總是會降低，在本文中，我們在執行完動態時間歸正後，以語音分段量測的方式來改善混淆群集的辦認，且僅僅只增加了一點點的運算量。在我們的實驗中，我們所用的語音資料是國語0~9的數字，在這10個數字中，含有幾個混淆群集，實驗結果示我們的方法確實有助於改善混淆群集的辦認。 The performance of DTW algorithm always degrades while applied on a vocabulary with confusion sets such as English alphadigits or Mandarin CV monosyllables. In this work, the DTW algorithm is argument with phonetic segmentation function to improve its discriminative ability only with a slight increase in its computation al load. The algorithm is tested on Mandarin isolated digit databases which contain the illustrative confusion sets. Simulation results show that it will be promising for recognizing the confusion sets in Mandarin speech recognition.