Enhanced Model of Long-Short Term Memory for Music Generation in Hardware

Abstract

Music generation becomes an economically important research field for the time being since the rapid growth of the entertainment industry. Among various methods for creating music, Long-Short Term Memory, or LSTM, is a preferable way for creating sequence of music notes while maintaining the harmony. This study proposes an enhanced LSTM model in hardware where the range of notes are also included in the generation process as a new factor. Data that comes in and out the LSTM are handled under normalization to reduce the range of note sequence, as well as to increase the consonance. Measurements were taken on the proposed LTSM, and another basic one, which were both implemented on hardware, to analyze the quality of the proposed LSTM and also the impact of the note range over the generation process. It appeared that the proposed LSTM could reach the harmony much efficiently based on the note range analysis; and achieve it faster than the basic LSTM since a smaller number of epoch was required. Those results indicate that LSTM can work with additional factor to improve its quality, rather than concentrating only on the data values.