即時單字音辨識系統之設計

Abstract

本論文之重點在於即時語單音辨識系統的設計與應用，以傳統的馬可夫模型為基礎；再結合擁有學習功能的向量量化演算法，使整個系統不但提高了辨識率並且計算量也大為減少；進而縮短辨識所需的反應時間，以符合即時語音辨識系統的需求。而在語音學習方面，我們將辨識錯誤的聲音直接與線上資料庫作修正，以快速並有效的方式達到我們所要的需求，以提高其辨識率。對於語音的訓練，針對未特定語者的語音辨識，我們使用維特比演算法來找出最佳的隱藏狀態序列，並根據這個序列上的狀態，利用k-means演算法來作分群，得到新的模型參數，將原本眾多的模型參數減少到一定有效的量，以符合所需。 除了單音辨識外，即時連續數字辨識也將大略的介紹，我們是利用以動態時間較準的方式為基礎的一階狀態演算法來求得連續語音的連接點，進而求得正確的辨識結果。利用一階演算法來作連續數字辨識，其資料庫的模型參數將會大大的影響其辨識結果，故其要慎選其訓練語料。

The major concept in this thesis is about the design of on-line isolated words recognition systems. By combining with the traditional hidden Markov model and learning vector quantization, not only the recognition accuracy will increase but also computations will decrease. Therefore, the reaction time of the speech recognition will also decrease. By using the learning function to adjust directly the on-line database with the misclassified patterns, it will be fast and useful to improve the recognition accuracy. For the data training, the Viterbi algorithm has been used to find the best stat sequence for speaker-independent. And the k-means algorithm has been also used to cluster the mean vectors and variance vectors in each state in order to decrease the number of models in our database. Besides, the connected digits recognition will be introduced conceptually. The one-state algorithm based on the dynamic time warping is used to recognize the connected digits. However, the models of the database will influence greatly the recognition result by using the one-state algorithm. So the training data must be chosen carefully.