Characterizing acoustic environments with OLAF and ELSA

Abstract

The confluence of signal processing and machine learning has created many innovative technologies in popular research areas such as speech recognition. However, many of the most successful methods are difficult to apply in areas that are not as popular and lack institutional support for research and creation of labeled data corpora. For instance, the expertise, computational resources, and quantities of data required for deep learning create a significant barrier to its implementation in many fields that it might otherwise impact. The focus of this work is the development of signal processing and machine learning methods that can be practically implemented with less human effort, less need for large quantities of labeled data, and less computational cost. Toward this goal, we have developed methods for outlier learning using augmented frozen dictionaries (OLAF) and estimating the likelihood of sparse approximations (ELSA) in the context of monitoring acoustic environments. Both methods utilize sparse, dictionary-based representations to capture information about the structure of the data. These methods are potentially applicable in many different areas and to many different types of data, but in this work have been tested specifically for monitoring poultry production facilities. The high levels of noise and uncontrolled nature of these environments pose significant challenges for audio signal processing efforts, especially when combined with a lack of labeled data. Our methods have proven effective at leveraging randomly selected unlabeled data and weakly labeled data to characterize the environment and highlight events or changes in the conditions in poultry houses. Providing these types of tools for monitoring livestock could help producers to better understand the effects different practices have on the animals and lead to better animal well-being.