基於雜訊模型之雜訊語音辨識技術設計

Abstract

要得到一個較準確的雜訊語音模型，最有效的方法就是收集足夠多的雜訊語音，經過重新訓練而得到一個匹配雜訊語音模型。然而，在實際運作的系統上，這種做法往往因為需要的資料量過大及需要花費過多的時間而窒礙難行。基於這兩個理由，本論文提出一個折衷的辦法。首先建立一個雜訊語音的資料庫。之後在進行雜訊語音辨識時，只需分析當時環境雜訊並與儲存於資料庫中的雜訊比對，找出最適合當時環境的雜訊語音模型，以此雜訊模型來進行辨識。如此一來，實際運作時就可以不用花費太多的時間而得到較準確的雜訊語音模型來進行辨識。本論文建立了四個雜訊語音模型的資料庫，並利用一個不在此資料庫的雜訊語音進行測試，在訊號對雜訊比為10dB時，句子辨識率可由1.25%提高至48.75%。

The best method to do the noisy speech recognition is to use the matched noisy models. However, in applications, it is difficult to re-train a noisy model because it requires large noisy speech data and a large amount of time. Therefore, an eclectic method is proposed in this thesis. The noisy speech models will be established off-line. While doing the noisy speech recognition, the recognizer will analyze the current background noise first, and then compare it with all the noisy models in the database to find a suitable noisy model. Finally, the noisy recognition will apply the selected noisy model. In this way, it could obtain a more precise noisy model without plenty of time. In this thesis, a database with four noisy models is established. The clean speech corrupted by noise, which is not in the database, will be recognized. The sentence correct rate can be raised from 1.25% to 48.75% in 10dB SNR.