國語聲訊處理

Abstract

在本論文中﹐我們針對語音與樂音之特性做處理﹐研究的主題可分為兩個 部 份來看﹕在第一部份﹐藉由頻譜訊息參數及韻律訊息參數之改變﹐對 中文文 句翻語音系統做語者之更換﹐其中頻譜訊息參數是從新語者錄製 之語料﹐抽 取國語語音411個音節波形樣板並偵測基週標位﹐而韻律訊息 參數是自新語 料中統計代表韻律訊息之聲母和韻母長度､停頓點長度及 可代表聲調之基週 軌跡參數﹐以調整韻律訊息之產生。第二個部份是做 聲訊處理﹐以產生特殊 音效﹐包括利用基週標位調整做語音聲調之升降 和播放速度之改變﹐用再取 樣法則做歌聲聲調之改變和放音速度之緩急 ﹐利用聲音延遲產生回音﹐設計 一頻譜轉換器做聲音轉換 In this thesis, several topics of signal processing for Mandarin speech and singing signal are studied. First, the technique of processing some utterances of a female speaker in order to add her voice to a Mandarin TTS system is studied. In this work, we first automatically segment all utterances into syllable segments, and then manually extract waveform templates of 411 synthesis units of base-syllable. Some prosodic parameters are then extracted and used to calculate their first- and second-order statistics in order to adapt the prosodic information synthesis to the style of the new speaker. Second, an LP model-based speech processing technique is proposed. The input speech signal is processed to shift the key, to change the speed, to simulate the echo effect, and to make a spectral transform. Last, a time-domain processing scheme for singing signal is discussed. The input singing signal is processed to shift the key, to change the speed, and to simulate the echo effect. Informal listening tests confirmed that all these proposed methods function well