使用可適性濾波器方法消除手機錄音時的背景聲音

Abstract

目前手機上有許多卡拉OK的APP，主要的功能是讓使用者選擇喜歡的歌曲伴奏並錄製歌唱。其中多數的APP有一項功能是歌唱評分，評分標準是根據使用者唱歌的音準。然而因為背景伴奏的干擾下，這類評分往往不夠準確， 本論文的目的是希望能透過可適性濾波器 (Adaptive Filtering)的方法來消除背景伴奏的干擾，以達到更加準確的歌聲評分。 本論文包含兩個實驗。第一項實驗是將手機錄音過程看為一個線性非時變系統(Linear, Time-Invariant System)，並找出一個合適的濾波器階數(Filter Order)，此濾波器是用來模擬系統的脈衝響應(Impulse Response)。第一項實驗的資料是將手機放在四種不同環境下的錄音資料，在床上，書桌上，浴室裡，以及拿在手上。計算濾波器的方法是使用最小平方法(Least Square Estimation)。 第二項實驗使用了四個可適性濾波器的方法，將卡拉OK錄音中的背景伴奏消除之後，再使用音高偵測(Pitch Tracking)的方法來評分人聲。最後會比較這四個可適性濾波器方法所得出的音高偵測結果以及演算法的計算速度。

There are many kind of karaoke apps on mobile devices. One feature of these karaoke apps is the functionality of scoring a user's singing based on pitch tracking. However, the scoring might be inaccurate if the singing voice is mixed together with the background audio. The objective of this thesis is to evaluate some adaptive filtering methods and try to subtract the background audio from the mixture signal to extract vocal part in order to enhance the accuracy of pitch tracking. The are two types of experiment in this thesis. First, modeling the mobile device's recording process as a linear, time-invariant system, a suitable order of the system's impulse response is chosen based on the first experiment result. In the first experiment, recordings are taken under four different environment, bathroom, bed, desktop, and holding by hands, and then a filter with suitable order to approximate to the system's impulse response is calculated by solving least squares problem. In the second experiment, four adaptive filtering methods are implemented and applied on some karaoke recordings to subtract the background audio, and then the resulting vocal part is processed for the pitch tracking using auto-correlation method. The pitch tracking results are compared with ground truth data for accuracy. Finally, four adaptive filtering methods are evaluated and compared in terms of pitch tracking accuracy and time complexity.