Transcription of Indian Classical Music using Convolutional Recurrent Neural Network and CTC Loss

Swaras (notes) combined with appropriate Alankars (ornamentations) play a vital role in evoking the aesthetic attributes of a Raag in Indian classical music. It adds to the complexity of recognizing Swaras present in a given composition. However, recognizing the employed Swaras is critical for determining many qualities of a Raag, although the process takes time owing to improvisations or ornamentations. In this study, we have curated a sizable transcription dataset containing 25-second segments of music files and their corresponding transcription containing annotations for Swara, Alankar, and silence or noise. We used two features: STFT, CQT, and tonic normalized versions of STFT and CQT, along with melody contour from a music file as input to autoregressive convolution neural network models. All the variants of models are trained using CTC loss. The experiments show that combining convolution layers with recurrent layers using the tonic normalized CQT and melody contour achieves a good performance of 13 % WER.