Forensics and biometrics
Using signal processing methods and technologies in human physiology.
Forensics and Biometrics (FaB) is concerned with advancing the state of the art in signal processing methods and technologies as they are used in the fields of forensic and biometric analysis.
Forensic and biometric signal processing are related, in that both fields attempt to infer information about an underlying subject based on measured, but incomplete, data samples. Our specific areas of academic interest include image and acoustic forensics, image and acoustic biometrics, acoustic analysis as it pertains to human physiology and numerous tangential applications of these methods.
Human biometrics is the automated recognition of a person using inherent, distinctive physiological or involuntary behavioural features. Example biometrics include speech, lips, fingerprints, irises, and gait. Our goal is to develop new algorithms to process unique biometric features extracted from different body parts.
Given the limitations of using a single biometric measure, we are attempting to improve the robustness and reliability of recognition by fusing multiple biometric measurements. For example, a hand is placed over a suitable sensor and features such as vein pattern and palm geometry are extracted for classification purposes. These two biometric features are then fused either in the feature domain or the classification domain to improve recognition performance.
Biometric-enabled systems are very much application-dependent so we need to investigate and develop expertise in many biometrics to best match the set of biometrics to a given application. For health applications, we need to consider non-contact sensors for hygienic purposes. Speech and palm vein patterns are ideal biometrics that comply with the non-contact requirement.
A long-term goal is the discovery and development of new biometric features.
Forensic Image Processing via Hyperspectral Methods
Hyperspectral images differ from regular pictures in that each pixel contains a spectrum of light over some portion of the optical spectrum. Typically that includes the visible and near infrared spectrum from approximately 400 nm to 1200 nm. One of our goals is to exploit the extra information available in hyperspectral images with a long-term plan to develop prototype tools to aid law enforcement professionals when examining a crime scene.
For example, we are developing the components of a hyperspectral crime scene camera that can identify candidate substances of possible forensic interest based on the spectral reflectance properties of the scene. The aim is for this to be achieved without contact with any material at the scene. This sort of work requires theoretical advances in algorithm development and signal representation as well as innovative hardware designs to make future devices portable and useful in the field.
Parallel research projects using hyperspectral methods include the development of highly robust skin detection methods for occupational safety; an iris recognition system integrating spectral and spatial information; and using multispectral data to aid in the automated processing of fresh produce.
We are actively looking for new applications of hyperspectral image processing techniques.
Modelling Speech Production
We are currently modelling the impact that aging has on speech production. Our studies focus on both the speech signal and models of the speech production system. To date our major focus has been on the vocal tract and the related acoustics. We have an acoustic reflectometer to measure vocal tract shapes and have obtained precise MRI measurements of the vocal tract shapes. From these shapes we can derive the resulting acoustic signal. The models can be improved with knowledge of the speech source and recently we have acquired a Laryngograph to measure glottal pulses. Applications of this work include speaker identification, speech recognition, speech synthesis, speech forensics (see below) and vocal health.
We also are developing synthetic voices for robots and have recently developed a synthetic New Zealand English Voice.
For many years the courts have made extensive use of recordings of conversations of offenders and suspects. Frequently these are covertly acquired via hidden listening devices or by eavesdropping on telephone and mobile communications. One common feature of such recordings is their very poor recording quality which makes subsequent determination of both content and speaker identity very challenging.
We have expertise in forensic speech analysis and are currently investigating the impact of cell phone technology on the acoustic parameters associated with speech that are important to the task of speaker identification in forensic applications.
General Speech and Audio Processing
We are also active in processing and modelling speech and audio signals for several applications such as audio watermarking, noise reduction and echo cancellation, speech recognition, speaker localisation, active noise control, automatic music assessment, music processing, and speaker recognition. A recently developed audio watermarking system is ready for commercial deployment.