iHear - Lightweight machine learning engine with context aware audio recognition model

With the increasing popularity and affordability of smartphones, there is a high demand to add machine-learning engines to smartphones. However, machine learning with smartphones is typically not feasible due to the heavy loaded computation required for processing large-scale data with machine learning. The conventional machine learning systems do not naturally or efficiently support some very important features for large-scale stream data.;To overcome these limitations, we propose the iHear engine that aims to support lightweight machine learning through a collaboration between cloud and smartphones. The contributions of this thesis are summarized as follows: 1) The iHear system architecture for achieving high performance with parallel and distributed learning by separating cloud-based learning from smartphone-based recognition 2) The context-aware model for improvement of the accuracy and efficiency in audio recognition and sound enhancement 3) Audio recognition with real-time data preserving data consistency 4) An intelligent hearing app for iOS devices developed for effective and dynamic audio recognition and enhancement depending upon users' context for providing better hearing experiences.;The efficiency and effectiveness of the iHear engine in terms of its continuous learning capability were evaluated on Apache Spark (MLLib) with audio recognition and filtering of streaming data.;We conducted experiments with multiple contexts of households traffic, offices, emergencies, and nature with real data collected from smartphones. Our experimental results show that the proposed framework for lightweight machine learning with the context aware model are very effective and efficient in terms of real time processing with a high accuracy rate of 90%, which is 20% higher than traditional approaches.