Open-Source Pipeline for Noise-Resilient Voice Data Preparation

Abstract

A crucial component for developing an automated speaker or speech recognition system is voice preprocessing, which filters unwanted noise and detects the speech part. This study aimed to develop a computerized model to remove background noise and improve signal quality for future applications. This model was developed on a large longitudinal dataset containing varying real-world noise and validated on both a public dataset and a locally collected test dataset in different environments. An overall voice pre-processing pipeline is presented in this study, including denoising, segmentation, feature extraction, and ease of storage. The backbone of the denoising model is a Kalman filter, where the parameters were obtained from a grid search method; the signal-to-noise ratio (SNR) was used as the performance metric. Also, the segmentation was done to remove the pauses from the audio signal before the feature extraction. Finally, the data was stored as a template with the most common features. The SNR result suggested that the Kalman filter-based proposed method performed successfully across diverse datasets. Thus, this model provides a robust and adaptable solution for real-world scenarios and also ensures the data storing quality for future applications.