林正甫
出 生 地
台北市
學   歷
桃園市立陽明高中
元智大學資訊工程系
國立台灣科技大學資訊工程所
興   趣
電玩,閱讀,聽音樂
生活剪影
- 指導教授:古鴻炎老師
- 中文題目:使用ANN抖音參數模型之國語歌聲合成
- 英文題目:Mandarin Singing Voice Synthesis Using ANN Vibrato Parameter Models
- 中文摘要:本論文針對歌聲表情的一個重要因素“抖音”,研究以短時傅利葉轉換和解析信號之方法來對歌聲音節作分析,而求得抖音參數。此外我們也將這個分析方法,應用於求取波形包絡的振動參數。求得各個音節的抖音和振動參數之後,再拿去訓練各項參數分別的類神經網路(artificial neural network, ANN)模型,之後依據所建造的ANN模型的輸出,再配合滿度、下拍點等規則,去控制諧波加噪音信號模型 (HNM)作歌聲信號的合成。經由主觀的自然度聽測實驗,所得的評分顯示,同時使用抖音和振動參數合成出的歌聲信號,的確可以比原始使用HNM合成出的歌聲信號有顯著的改進。
- 英文摘要:In this thesis, analysis and synthesis of vibrato, an important factor of singing expression, are focused. We analyze the vibrato parameters of a singing syllable by using short-time Fourier transform and the method of analytic signal. In addition, we apply the same procedure to analyze the vibrating parameters from a syllable’s waveform envelope curve. When the parameter values of vibrato and amplitude vibrating are obtained for each singing syllable, they are used to train an artificial neural network (ANN) based model for each different parameter type. Then, these ANN models are used to generate the vibrato and vibrating parameters. Next, these parameters and other relevant music parameters are used together to control a harmonic-plus-noise (HNM) model to synthesize singing voice signals. With the synthetic singing voices, subjective perception tests are conducted. The result show that the singing signal synthesized with the control of vibrato and vibrating parameters is indeed apparently better than the singing signal synthesized without such controls.