摘要

    本論文提出一個多頻帶線性預估倒頻譜係數(multi-band linear predictive cepstral coefficients)的語音情緒特徵，利用離散小波轉換將訊號分解至多個子頻帶，對全頻帶和每個頻帶萃取出線性預估編碼係數，同時分析不同參數多頻帶線性預估倒頻譜係數，最後決定以分解2層、10階線性預估編碼係數和縮短取樣比例為8的做為參數。並且結合音高和能量曲線特徵，總共有52特徵，最後藉由費雪比例選擇出32個做為7種情緒的語音情緒辨識系統特徵，其整體辨識率達到90%。
最後本論文比較三種不同的類神經網路辨識器(多層感知機、放射基底函數網路和多維矩形複合式神經網路)。在整體資料集辨識率，多層感知機有90% 以上的最佳辨識率；模糊化多維矩形複合式神經網路對於訓練資料有著高達百分百的辨識結果；最後放射基底函數網路在測詴資料集有68% 的辨識率。

關鍵字：語音情緒辨識、音高、多頻帶線性預估倒頻譜係數、費雪比例、類神經網路、模糊化多維矩形複合式神經網路

The Analysis and Recognition of Emotional Speech Using Artificial Neural Networks

ABSTRACT

    This thesis presents a multi-band linear predictive cepstral coefficients (MBLPCC) feature for the emotional speech recognition system. Base on discrete wavelet transform (DWT), the emotional speech is decomposed into various frequency subband, and LPCC of the lower frequency subband for each decomposition process are calculated.
Furthermore, we analyze the different parameters of MBLPCC, and then decide to decompose two times, 10 LPCC coefficients and the downsampling ratio of eight as the parameters. We also combine MBLPCC with pitch and energy curve features, a total of 52 features, and choose 32 features by Fisher‟s ratio for the seven kinds of emotion of emotional speech recognition system, and achieves the recognition rate of 68%.
Finally, we compare three different artificial neural networks (ANN) recognizer, multilayer perceptrons (MLP), radial basis function networks (RBF) and fuzzy hyperrectangular composite neutral networks (FHRCNN). In the recognition rate of overall data set, MLP achieved the best rate of over 90%. FHRCNN with training data set achieves recognition result of up to 100%. Finally, RBFN with testing data set achieves the recognition rate of 68%.

Keyword: emotional speech recognition, pitch, MBLPCC, Fisher‟s ratio, artificial neural networks, FHRCNN.