The objective of the Acoustic Analytics: Acoustic Event Detection, Classification and Analysis research project is to develop key signal processing and analytical tools to extract timely, useful, and actionable information from real-world acoustic data on a large scale. Specifically, we are aiming to build acoustic analytic systems for 24/7/365 machine-automated monitoring of human environments on a large scale. The systems should be general enough, fast enough, and robust enough to yield useful information from large volumes of acoustic data.
Sound is second only to vision as a means by which humans sense and understand the world. The severity of deafness as a disability reflects concisely the importance of sound in understanding what’s happening in the physical world, including omni-directional and out-of-sight awareness of events and warning of danger. We thus believe that acoustic sensing, sense-making, and analytics of audio data could prove as significant to machine-automated monitoring of human environments as it is to the humans. The massive deployment of mobile phones and microphones in most personal computing devices has raised the quality and driven down the cost of acoustic data acquisition immensely over the past two decades, as well as provided the capacity to form and deploy massive, city-wide networked arrays of acoustic sensors. Smart-phone technologies enable significant computation for real-time point-of-acquisition data analysis. Thus in the past few years, real-time 24/7/365 acoustic monitoring on a very large scale has gone from inconceivably expensive to eminently feasible. In spite of this potential, acoustic data analysis (other than speech recognition) has been much less studied and deployed in comparison with vision.
Today, acoustic data analytics are based on techniques tailored for specific applications (SONAR detection of submarines; automatic speech recognition for specific languages and contexts; detection of whale calls of a specific species; gunshot detection) that do not generalize effectively. These techniques are often very expensive computationally and so are unsuitable for 24/7/365 analysis of large volumes of audio data streams. Therefore, advances in audio signal processing and acoustical analytics are required to process large volumes of audio data to extract timely, useful, and actionable information from the real world.
Audio-based monitoring also has significant impact on surveillance for public safety and security and on urban noise assessment of residential area. Potential safety hazard is raised for crowded public areas like drinking bars and the places for recreation and leisure, especially during nighttime. Crimes like robbery also happen in hidden public areas without awareness from neighbors. It is crucial to detect such emergency cases in a timely manner and alert police for preventing further damage. Audio-based monitoring systems have proven to be very useful tools for detecting such cases. It might normally be the case that there is shouting/screaming/crying sound from these kinds of incidents. Timely acoustic sensing and correctly detecting the anomaly events become very important.
Likewise, urban noise effects are becoming more and more serious to human health. Traffic, business and even recreational activities all contribute to spoiling a city and harming its inhabitants by exposing them to undue levels of noise. Noise issues have to be carefully analyzed and controlled. Noise mapping and prediction is an essential tool to aid the assessment of noise levels over a wide area and to predict the changes in the noise environment due to changes in use. Creating an accurate noise map will be very useful in communicating issues and defining future policy, such as to communicate the noise situation to stakeholders, to inform areas of planning such as construction, traffic & transport and to build a common understanding within the community.
Towards our vision of developing the capability to extract timely, useful, and actionable information from real-world acoustic data on a large scale, this project is divided into several major tasks:
- Acoustic event detection: performing universal acoustic event detection for quickly discarding the vast majority of uninteresting data in the analysis of massive audio datasets, and improving essential robustness to adapt to different background noises while retaining high sensitivity to detection events;
- Acoustic event classification: classifying the detected events into different classes that are precisely defined for human understanding with improved robustness to environmental noise and interference;
- Anomaly detection: detecting rare or never-before-observed unusual acoustic events;
- Acoustic energy density mapping: monitoring environmental noise levels and mapping over space and time the intensity and other parameters of the noise field;
- Real-time audio analytics: performing high-level audio analysis to exploit and extract the known structure in acoustic data for further data analytics and mining.
- Sonification: encoding of non-auditory information into sound for “visualization” for increasing the information-delivery capacity and representing certain types of data more effectively.