基於時域上基週同步疊加法之歌聲合成系統

Singing Voice Synthesis System Based on Time Domain-Pitch Synchronized Overlap and Add

在本研究中，我們提出並實作一個串接式的歌聲合成系統，用來產生具有配樂的合成歌聲。語料庫的錄製是根據注音符號檢字表來錄製，並錄製三種不同的音高。我們將主旋律中的力度、音符編號、起始時間和結束時間來當作合成資訊，並加入了轉音的資訊。在合成單元的處理上，採用時域上基週同步疊加法來對合成單元做時域上的修改。我們提供一個歌曲的選擇介面供使用者來進行歌曲的合成，並加入了一些對於合成歌曲的調整。包括了整體上音符編號的調整、歌詞的修改等等。此外，也做了一些聽測實驗，來進行合成歌曲的品質、清晰度和相似度的評估。品質評估方面，合成歌曲加上配樂有改善的效果。清晰度和相似度評估方面，簡單的歌曲有較好的表現。

In this study, we propose and implement a concatenation-based singing synthesis system to synthesize the singing voice with background music. We record three di erent pitches to build our corpus for all syllables. The synthesis informations, including velocity, note number, start time and end time are extracted from the main melody. Runs and ri s information was added into consideration afterward. We use TD-PSOLA to modify the synthesis units in time domain. At last, we add back the background music extracted from MIDI to our synthesis song. We implemented a user interface for users to synthesize songs. This interface can be used to adjust the synthesis songs, for example, adjust the overall pitches in the song, modify syllables, etc. Finally, we evaluate the quality, clarity and similarity of the synthesis songs. The results show that the proposed method achieve better results with simple songs than with fast songs.