使用自我組織圖建構的音樂資訊記憶模型

Abstract

在電腦音樂的領域裡，人們嘗試各種方法來讓電腦辨識音樂，理解音樂中的音高，音色，和弦，樂曲風格等等。大多數的電腦聆聽系統使用統計方法，或是數位信號處理。統計方法及數位信號處理功能強大具有實用性，但是在也有存在些限制，因此我們嘗試模擬人類記憶的特性作為辨識的基礎。在這篇論文中，作者嘗試建構一個關連式記憶體，利用記憶的關連性做為音樂辨識的基礎。 本論文，使用類神經網路技術來建構關連式記憶體。關於關連式記憶在類神經網路中已經有過許多討論以及實做，像是Hopfield Network與改良式Hopfield Network或是各種Recurrent Neuron Network等等。在此我們使用SOM來建構我們的記憶模型來記憶音樂資訊。相對於其他種類的網路，使用SOM建構的記憶模型對於空間與時間的複雜度需求較低，能夠記憶依時間順序變化的資料，適合用來處理音樂資訊。

In the field of computer music, people have tried many methods to let computers recognize and understand music. For examples, Programs for recognizing pitch, tone, chord, and genre of music have been developed. Most of such music recognition or music-listening systems, people used statistics or digital signal processing to solve this kind of problem. Although these methods are powerful and useful, they have their own limits. In this thesis we try to simulate human memory for recognizing music. We construct an associative memory system, and use the association to recognize music. Our associative memory system is based on neural network. Associative memory is not a new topic in neural network research. There have been many papers about associative memory since 1989. The major categories of associative memory include Hopfield model, Hopfield-like model, mind-in-a-box, and SOM based associative memory. Here we use SOM as a supervised sequential learning system for memorizing music information. With this system, the input midi music data passes the pre-processing and the training process, to generate memory. The recall will be conducted by pieces of music.