哼唱式音樂檢索系統之容錯能力提升

Abstract

隨著數位音樂數目的成長，如何正確且及時地從音樂資料庫中搜尋出特定的歌曲資訊，實為一項重要的課題。哼唱式音樂檢索(Query-by-Singing/Humming，簡稱QbSH)不同於以往利用文字為基礎的搜尋方式，其透過使用者哼唱歌曲的部分片段，從中計算隨時間變化的音高(Pitch)特徵，並與資料庫進行比對，以輸出最符合使用者哼唱的歌曲資訊。現今已有許多QbSH的相關研究和應用軟體，且對於使用者正確哼唱皆能提供良好之比對結果，然而，倘若使用者哼唱時受雜訊干擾，或其自身在停頓、音準、以及節奏出現錯誤，則目前的研究和軟體將難以處理上述多種錯誤情況。本研究提出一具容錯機制之QbSH系統，其透過系統架構和演算法設計，實現雜訊濾除、停頓處理、音高調整、以及節奏處理之容錯能力。本研究之比對結果，並與目前行動裝置中常用的QbSH應用軟體作比較，以驗證其優越性。

With the proliferation of digital music, there is an important topic on searching the accurate and specific song from music database and in a proper real-time. Query-by-Singing/Humming (QbSH) is a song search technique that differs from previous text-based search. It compute the time-varying pitch feature from user’s acoustic input, and compare with music database to generate the ranking list of similar songs. Related QbSH researches provide the good matched results only when user sang correctly. If user was interfered with noise when singing, or user occurs the rest, pitch, or tempo mistakes, current researches may not handles above conditions. A novel QbSH system with the design of new system architecture and algorithm has been proposed to achieve the fault-tolerance of noise, rest, pitch, and tempo. The experimental results achieved the superiority by comparing to current QbSH software.