结合多通道处理的智能音符起点检测方法

音符起点检测似中的一个棘手问题是:检测门限不仅因乐曲的不同而不同,而且对同一首乐曲的不同段落也不一样。此前提出的基于音符平均能量NAE(NoteAverageEnergy)的时域方法,虽然摆脱了门限选择问题,但它要求乐曲功率络的音符模式,要么是硬模式(快起慢落)要么是软模式(慢起快落)。智能检测方法首先对整首乐曲功率包络的变化按不同模式划分为若干段落,然后对不同的段落施以不同的检测准则,这使它能胜任更加复杂的混合模式。在成功识别音符模式的基础上,漏检音符的查找策略使智能检测方法的检测率显著提高。对各种乐器和曲风的乐曲所做的大量实验表明:智能检测方法能够准确检测出80%以上的音符。结合了多通道处理技术的智能检测方法,使检测率又提高了10%。

Smart music onset detection combined with band-wise processing

In the music onset detection, there was a thorny unexpected problem: the threshold for detection can not be fixed for all songs, neither for the different segments of one song. The note average energy (NAE) based time domain approach proposed earlier gets rid of the threshold determination, but it requires the note pattern in the power envelope of the song to be either hard (sharp raise smooth fall) or soft (smooth raise sharp fall). The smart detection approach developed in this paper makes it capable for more complicated pattern situation: It marks the pattern variation of the power envelope of the whole song first, and then applies different detection criteria to segments with different pattern mark. With the great success in note pattern recognition, a missing-note-check-back strategy is introduced to the core of NAE based smart onset detection. Over 80 percent of the notes can be detected based on our experiments of wildly chosen songs of different instruments and performance styles. Combining band-wise processing, detection rate of the smart detection can be improved by extra 10 percent.