語音強化技術在相加性雜訊環境下的語音辨識之研究

Abstract

環境雜訊的干擾是導致目前語音辨識技術無法普遍應用在實際環境中的瓶頸。爲此，本論文針對了相加性雜訊環境下的語音辨識系統，提出了強化型MMSE語音強化法，以消除環境雜訊對語音的干擾。此方法是以最小平方誤差短時頻譜振幅估計法為基礎，並考慮語音訊號與雜訊訊號在某段時間中的變動程度，去調整濾波器的頻率響應，以達到強調語音訊號並壓抑雜訊訊號的目的。 我們根據AURORA提出的語音辨識架構進行實驗。實驗結果說明了：1. 透過根據時間變動程度的調整方式，強化型MMSE語音強化法確實能夠增加強化後語音特徵中差量參數的正確性；2.與其他的語音強化法進行比較，本方法也能夠在準確率上有所提升。我們並將此方法實作在一個分散式語音辨識系統上，經由多位使用者實際操作後，確實能有不錯的辨識效能。

In practical environment, the speech recognition performance degrades drastically due to the background noise interference. For this reason, we propose the enhanced MMSE speech enhancement approach for the speech recognition in additive noise environment. This approach is based on Minimum Mean-Square Error Short-Time Spectral Amplitude Estimator, and adjusts the filter frequency response according to the variation of speech and noise in the local time period, in order to boost the speech variance and suppress the noise variance. The experiment follows the AURORA proposed architecture. The result shows this adjusting approach increases the correctness of delta-coefficient, and has better accuracy comparing to other speech enhancement method. Moreover, we apply the proposed method and implement a distributed speech recognition (DSR) system.