Indian Music Instrument Classification Using Deep Learning on Embedded Platforms

Indian musical instruments are an integral part of the rich cultural heritage of India. These instruments have evolved over centuries and are unique in their sound, construction, and playing technique. The popularity of Indian music in Western countries has given rise to the need of identifying Indian musical instruments. In this study, we aim to classify Indian musical instruments by extracting Mel-Frequency Cepstral Coefficients (MFCCs) from the audio signals and training the Recurrent Neural Networks’ (RNNs) architecture on these features. We compare the performance of the RNN with various Machine Learning (ML) techniques and find out that RNN outperforms all the other techniques. We quantize the trained RNN model to decrease latency by 13% and model size by 87.2% with negligible change in the model inference accuracy compared to the baseline model and deploy it on an Android device making a user-friendly application. This application allows users to record audio clips and classifies the audio into five different Indian musical instruments and other sounds as background noise. The app helps users to recognize Indian musical instruments in a convenient manner.