Music Note Series Precipitation using Two Stacked Deep Long Short Term Memory Model

Abstract

People need to get relieve from their stress and thoughts by engaging themselves with entertainment. Music plays a vital role in changing the people environment to overcome their personal problems. As technology is involved in all the fields, deep learning is extensively contributing its performance in the music industry towards generation of music note sequence. Music confirms to be a tough dynamic than image data as it is temporal with hierarchical structure and cross-temporal dependencies. As music is composed of multiple instruments by being interdependent and being evolved over time, it remains a challenging issue for the researchers to work. Since music is distributed into chords, arpeggios, melodies and each time-step generating multiple outputs, the generation of music note sequences through deep learning network extend higher issue for the researchers in programming. By interpreting the above scenario, a new two-stacked Deep Long Short Term Memory (TSD-LSTM) model is designed for music note sequence generation using audio framework of piano with 16000 frame length. The dataset is extracted from Maestro database repository. The audio framework data is converted into MIDI format using pretty_midi package. The MIDI digital data is parsed with 30513 number of notes resulting with sequence length of 25 frames and vocabulary size of 128 pitch length. The total digital data is ended up with 1282 MIDI information frames. The parsed audio digital data is splitted into training data with 967 MIDI data, validation data with 137 MIDI data and testing data with 178 MIDI data. Model fitting is done with TSD-LSTM. The proposed model design is refined with parameter optimization. The project is implemented with Python under NVidia Tesla V100 GPU server with 215938 trainable parameters, training epochs of 300, batch size of 64 along with training rate of 0.001. The music note sequence generation is done with TSD-LSTM and fitted with single stack LSTM, 3 stack LSTM and 4 stack LSTM and the performance is analyzed and compared with metrics like train loss, duration loss, pitch loss and step loss. Implementation results shows that TSD- LSTM model have minimum value for test loss of 0.0567, duration loss of 0.0086, pitch loss of 0.8150 and step loss of 0.0073 compared to other models.