LeaderSTeM-A LSTM model for dynamic leader identification within musical streams

PREPARING DATASET

We used the file 'GetFeaturesAubio.py' to get all the features from the STEM files of MUSBD18[1]. We used AUBIO [2] beattracking and pitch, amplitude extraction techniques to generate the realtime data sequence.

ARRANGING THE DATASET

ArrangeDataset.py - This file was used to arrange the values generated using 'GetFeaturesAubio.py'.

DATASET

Excel folder contains the dataset file we used.

There columns are 16 columns +------+-----------+-------------+--------------+---------+-----------+------------+---------+-----------+------------+--------------+----------------+-----------------+-----------+-------------+--------------+ | Time | OutputBPM | OutputPitch | OutputVolume | DrumBPM | DrumPitch | DrumVolume | BassBPM | BassPitch | BassVolume | AccompanyBPM | AccompanyPitch | AccompanyVolume | VocalsBPM | VocalsPitch | VocalsVolume | +------+-----------+-------------+--------------+---------+-----------+------------+---------+-----------+------------+--------------+----------------+-----------------+----------

MODELS

The codes can be found in the respective folders. 3-Layers - This folder contains six different Tuning Codes. All the models were 3-Layered LSTM.
4-Layers - This folder contains six different Tuning Codes. All the models were 4-Layered LSTM. 5-Layers - This folder contains six different Tuning Codes. All the models were 5-Layered LSTM.

Ray Tuner was used to find the optimal model. [3]

REFERENCES

[1] Rafii, Zafar, et al. "MUSDB18-a corpus for music separation." (2017). [2] https://aubio.org/ [3] https://docs.ray.io/en/master/tune.html

Presentation

https://www.youtube.com/watch?v=TOC4cn7jOGA&ab_channel=AIMusicCreativity2020