Music chord inversion shape identification with LSTM-RNN

The art of music production has changed slowly over time with the technological advances. Multifarious automated solutions have lent a helping and to the musicians in different ways from practice to production and stage performance. In the context of a musical composition, the background music (BGM) is extremely important as the lead melody. One of the building blocks of BGM is a chord which is composed of two or more musical notes played simultaneously. Each chord can be played in multiple ways which adds to the melodic variety. Each of these ways is termed as inversions whose identification is extremely important for analyzing compositions and transcribing them. It is also extremely important for automated BGM or lead melody generation, where the inversion form or shape of a chord plays a pivotal role in the feeling of a composition. The challenge of chord shape identification further increases for clips of shorter length which is very critical for real time processing. In this paper a system is presented which distinguishes chord shapes from clips of extremely short durations. Experiments were done with as many as 40572 clips recorded in ordinary room environment and a highest accuracy of 99.47% has been obtained with LSF-deltaS deltaG features and LSTM-RNN-based classification.