Identifying Musical Instruments with Machine Learning

Description:

The automatic identification of musical instruments is an important task in music information retrieval. Previous approaches often rely on spectral analysis to identify the frequency patterns needed to distinguish between instructions. In this work, we present an approach to automatically identify musical instruments directly from the raw audio waveform, without performing any spectral analysis or feature extraction. For this task, we trained a bidirectional LSTM, a type of recurrent neural network often used to learn sequential data. The model was trained on sequences of 256 bits from one-second samples of single-channel 16 kHz wav files. The model was trained on audio excerpts from the RWC Music Database, which contains examples of musical instruments played across their range and with different dynamics. To prepare for learning, the data was first split into individual, one-second chunks and downsampled to reduce the amount of data. The model was trained on ten instruments to perform multiclass classification directly from raw audio sequences.