SurShastra - Music Transcription Software

Abstract

SurShastra is an automatic music transcription (AMT) system designed to convert polyphonic audio recordings into symbolic musical notation, such as MIDI files. While AMT has made significant progress in Western music, accurately transcribing non-Western musical forms, particularly Indian classical music, remains challenging due to microtonal variations, complex rhythms, and improvisational elements. This project addresses these challenges by combining advanced deep learning methods, specifically Convolutional Neural Networks (CNNs) and Bidirectional Gated Recurrent Units (Bi-GRUs), to model both the spatial and temporal characteristics of musical audio. The system processes input audio into log-mel spectrograms, which are then used to predict note onset, offset, velocity, and frame-wise note presence. Unlike conventional frame-based models, SurShastra incorporates onset and offset regression to estimate the precise timing of musical notes, improving transcription resolution beyond frame-hop limitations. The model is trained on the MAESTRO dataset to ensure robust performance across different musical compositions and styles. Evaluation using precision, recall, and F1 score demonstrates the models ability to transcribe complex polyphonic pieces accurately. Preliminary results indicate promising performance in capturing musical nuances such as ornamentation and expressive timing, particularly in Indian classical music. The proposed system aims to bridge the gap between musical audio and symbolic notation, enhancing accessibility, analysis, and learning across both Western and non-Western musical traditions.