Feature for Musical Pitch Estimation from Simplified Auditory Model

A simplified auditory model has been used for calculating an enhanced summary auto-correlation or ESACF, which can be used as a tool for musical pitch estimation from audio signal. The model itself is not only computationally efficient but its ESACF also shows a good result for single pitch estimation. However, using this ESACF for multiple pitch estimation seems to be very difficult to analyse because musical instruments usually have timbre variations even for the same kinds of musical instruments. By modifying this model, we can generate input features to use with neural network for assisting the process of multiple pitch estimation. Thus, each output of the neural network is mapped to each musical pitch and used to indicate each existing pitch probability. In our experiments, we generated data sets from recording of real musical instruments and used these data sets to train neural network and evaluate its performance. We compare performances of neural network between using of these proposed features and spectral features generated from audio spectrum. From the results, we found that the performances from these proposed features can be comparable with the features generated from audio spectrum and some experiments illustrated that these features yield better performances for musical instrument signals with slightly changes in their timbres.