本論文對於哼唱選歌進行了加速與準確度的改良。我們提出了同時使用旋律和歌詞的資訊的方法。首先會進行哼唱分辨,以將「唱」和「哼」分離開來。對於「哼」的查詢,我們套用了只使用音高資訊的旋律辨識方法;對於「唱」的查詢,我們將旋律距離和歌詞相似度的結果合併,以利用額外的歌詞資訊。本論文中也使用了圖形處理器來進行加速旋律辨識的部分,我們選擇最耗時的資料庫比對部分來加速,並嘗試不同的平行方式以達效能最佳化。
This thesis proposes the acceleration and accuracy improvement of a query-by-singing/humming system. We use both melody and lyrics information to achieve better accuracy for query-by-singing/humming. Singing/humming discrimination is first performed to distinguish singing from humming queries. For a humming query, we apply a pitch-only melody recognition method. For a singing query, on the other hand, we combine melody distance and lyrics similarity to take the advantage of extra lyrics information. We also use graphical processing units to accelerate the melody recognition module. We choose to accelerate database comparison, the most time-consuming component of the system, and try different methods to optimize the performance.