Musical Instruments Classification Using Long Short-Term Memory Network

Chiah, Cai Xiu (2022) Musical Instruments Classification Using Long Short-Term Memory Network. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Chiah Cai Xiu.pdf Restricted to Registered users only Download (2MB)

Abstract

Computers are not able to recognize musical instruments from the audio files as human do by merely listening to them. However, when it comes to a large number of files required, human may not be able to classify each unknown audio file in certain time constraint. This project focused on an underrated deep learning technique, Long Short-Term Memory (LSTM) to recognize and classify different musical instruments from the audio signals. LSTM, as an advanced version of recurrent neural network (RNN), have an excellent performance in processing time-series data and complex data pattern learning. Hence, this project aimed to develop a bidirectional LSTM and to study the feasibility and accuracy of the proposed model in musical instruments classification with log Mel-spectrogram. With a dataset of 14 musical instruments, the experimental results demonstrated that the proposed model obtained a micro F1 score of 0.977 and 0.978 for macro F1 respectively, showing a good performance as the current state-of-art, CNN model on monophonic based recognition tasks.