音段式處理技術在中文語音處理之應用

Abstract

這篇論文研究有關音段式處理技術在中文基週軌跡合成、低非固定速率語 音頻譜編碼及音節辨認等問題的應用。在中文基週軌跡合成問題上，以一 固定維度之直交多項式展開係數來代表基週軌跡，對此係數以向量量化產 生碼書，以此碼書及由語句中所抽出之語言特徵參數共同訓練出一統計模 型，此模型具有描述由語言學家所分析歸納之基週軌跡合成規則的能力。 接著，提出以軌跡量化為基礎之低非固定速率語音頻譜編碼方法，對由預 切割所得之音段中每一頻譜參數軌跡的直交多項式展開係數直接進行量化 ，為了改進編碼效率，將切音、量化及碼書設計作最佳化過程。將音段式 特徵參數應用至中文語音單音節辨認問題上，有兩種不同的辨認方法被提 出，第一種辨認方法將語音信號中的每一音段以一音段式特徵參數來代表 ，第二種辨認方法為一修正式隱藏式半馬可夫模型，其每一個狀態的語音 信號以非穩態性的高斯分佈序列來描述。利用以音段式特徵參數之統計性 質所建立的音節邊界模型來配合音節隱藏式馬可夫模型，在處理連續音節 間連音效應的問題上也得到有效的成果。 In this dissertation, applying the segment-based processing technique to some Mandarin speech signal processing applications including fundamental frequency (F0) synthesis, low data rate variable-frame-rate coding of spectral information, and speech reccognition are studied. A statistical model is used to describe the relationship between the F0 contour patterns of syllables and linguistic features representing the context. A novel segment vocoder based on contour quantization is proposed. Two coding schemes based on different segmentation strategies are studied.One is a basic coding scheme which performs segmentation and quantization sequentially. The other is an optimal coding scheme which performs segmentations, quantization, and codebook updating simultaneously. Two segment-based Mandarin syllable recognition schemes are then proposed. One is a template matching approach using segmental feature vectors and the other is a modified hidden sime-Markov model. To compensate for the inter- syllable coarticulation effect on syllable-based continuous Mandarin speech recognition, some inter-syllable boundary models were trained and directly supplemented to a set of context- independent syllable CHMM models to help speech recognition.