使用能量補償之雜訊國語語音辨認

Abstract

本論文提出了一個新的能量補償的方法，利用訓練階段時所求得的 411音節的能量曲線(energy contour)及正規化的 參數，使其能在雜訊環境下測試時能動態追蹤目前音段的語音信號強度，再配合 PMC 作背景雜訊的補償，以及通道偏移量補償，期待經過如此補償的模型，能在不同的測試環境下都能有一定的穩定性是本論文的最終目標。

In this thesis, we implement an algorithm which utilizes the energy information embedded in Mandarin 411 syllables to attain reliable gain compensation. Compared to parallel model combination (PMC) method which estimates a rough gain for the entire utterance, our method may trace variation of the volume of each syllable dynamically. Besides this, some other ways to make the trained model more compact are added, like Wiener filtering to suppress the background noise, signal bias removal to eliminate the effect incurred by the communication channels. Ultimately, our goal here is to provide an environment-robust method under different testing conditions.