Improved Feature Extraction for Time Series Data Using Sliding Window: A Case Study of Carnatic Tala

Human action patterns are typically temporal. Since the data collected is periodic, processing time series data into a feature matrix for further learning and pattern recognition is a challenging task. Periodic data must be processed in a sequential manner. Here we have tested the case study to identify the patterns related to the Carnatic Sapta Tala (the Seven Tala System) used in Carnatic music. Accelerometer, gyroscope, gravity sensor, and rotation vector are some of the sensors embedded in smartphones and used for data collection. The minimum window size to extract distinct characteristics of the selected Sapta tala was determined to be 20 through the experiment performed. This study employed a feature extraction method with a 20-second window and 99.95% overlap to create a feature matrix for identifying Sapta tala patterns in the Carnatic music system. The findings demonstrate an improved association between these rhythmic patterns of movement and can be tracked using smartphone sensors when utilizing a method that involves 20-frame overlapping windows. Future work could incorporate more advanced machine learning algorithms and data from additional sensors to further enhance the proposed methodology to categorize tala patterns. The optimal window size of 20 frames can be further enhanced during the application of classification algorithms, subject to the relevant metric.