Artificial Intelligence Methods for Automatic Music Transcription using Isolated Notes in Real-Time

Abstract

We introduce a comparative study of several features obtained from audio signal and methods of Artificial Intelligence employed for Automatic Music Transcription in real-time, specially using monophonic notes. Mel-frequency Cepstrum Coefficients (MFCC), Linear Prediction Coefficients (LPC) and Cochlear Mechanics Cepstrum Coefficient (CMCC) were the features used which are a set of coefficients obtained from our laboratory experiments, which in this paper demonstrated to be more effective for Automatic Music Transcription (ATM) than other characteristics such as Mel Frequency Cepstral Coefficients (MFCC). At same time, Vector Quantization (VQ), Hidden Markov Models (HMM), Gaussian Mixtures Models (GMM) and Artificial Neural Networks (ANN) for pattern classification task were used. The database consisted of 840 music notes, we analyzed 5 scales and 14 samples by musical note. The results obtained showed that Vector Quatization, HMM using CMCC_L&B_RA and GMM were the best methods of Artificial Inteligent for this task, while MFCC and CMCC_L&B_RA were the best features employed.