Speech processing for forensic applications
Peter Barger · QUT ePrints (Queensland University of Technology) · 1998
This thesis examines speech processing systems appropriate for use in forensic analysis. The need for automatic speech processing systems for forensic use is justified by the increasing use of electronically recorded speech for communication. An automatic speaker identification and verification system is described which was tested on data gathered by the Queensland Police Force. Speaker identification using Gaussian mixture models (GMMs) is shown to be useful as an indicator of identity, but not sufficiently accurate to be used as the sole means of identification. It is shown that training GMMs on speech of one language and testing on speech of another language introduces significant bias into the results, which is unpredictable in its effects. This has implications for the performance of the system on subjects attempting to disguise their voices. Automatic gender identification systems are shown to be highly accurate, attaining 98% accuracy, even with very simple classifiers, and when tested on speech degraded by coding or reverberation. These gender gates are useful as initial classifiers in a larger speaker classification system and may even find independent use in a forensic environment. A dual microphone method of improving the performance of speaker identification systems in noisy environments is described. The method gives a significant improvement in log-likelihood scores when its output is used as input to a GMM. This implies that speaker identification tests may be improved in accuracy. A method of automatically assessing the quality of transmitted speech segments using a classification scheme is described. By classifying the difference between cepstral parameters describing the original speech and the transmitted speech, an estimate of the speech quality is obtained.